alembic/versions/0079_code_shapes_ledger.py

@@ -0,0 +1,66 @@
+"""The shape ledger: code_shapes (#2787, milestone 294)
+
+Revision ID: 0079
+Revises: 0078
+Create Date: 2026-08-19
+
+The accounting half of the pattern system (governing note 2786): the snippet
+library records canon (small); this table accounts for EVERY shape the
+coverage extractor finds in a bound repo (total). Rows arrive `unclassified`
+from the coverage sync (step 2) and gain judgments — canonical / instance /
+variant / exempt — from audits, hooks, and the mechanical proposer.
+Unclassified IS the todo list.
+"""
+import sqlalchemy as sa
+from alembic import op
+
+revision = "0079"
+down_revision = "0078"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    op.create_table(
+        "code_shapes",
+        sa.Column("id", sa.Integer(), primary_key=True),
+        sa.Column(
+            "project_id",
+            sa.Integer(),
+            sa.ForeignKey("projects.id", ondelete="CASCADE"),
+            nullable=False,
+        ),
+        sa.Column("repo_key", sa.Text(), nullable=False),
+        sa.Column("path", sa.Text(), nullable=False),
+        sa.Column("symbol", sa.Text(), nullable=False),
+        sa.Column("kind", sa.Text(), nullable=False),
+        sa.Column("status", sa.Text(), nullable=False, server_default="unclassified"),
+        sa.Column(
+            "snippet_id",
+            sa.BigInteger(),
+            sa.ForeignKey("notes.id", ondelete="SET NULL"),
+            nullable=True,
+        ),
+        sa.Column("reason", sa.Text(), nullable=True),
+        sa.Column("classified_by", sa.Text(), nullable=True),
+        sa.Column("classified_at", sa.DateTime(timezone=True), nullable=True),
+        sa.Column("first_seen_commit", sa.Text(), nullable=False, server_default=""),
+        sa.Column("last_seen_commit", sa.Text(), nullable=False, server_default=""),
+        sa.Column("vanished_at", sa.DateTime(timezone=True), nullable=True),
+        sa.Column("created_at", sa.DateTime(timezone=True), nullable=False),
+        sa.Column("updated_at", sa.DateTime(timezone=True), nullable=False),
+        sa.UniqueConstraint(
+            "project_id", "repo_key", "path", "symbol", "kind",
+            name="uq_code_shapes_identity",
+        ),
+    )
+    op.create_index(
+        "ix_code_shapes_project_status", "code_shapes", ["project_id", "status"]
+    )
+    op.create_index("ix_code_shapes_snippet", "code_shapes", ["snippet_id"])
+
+
+def downgrade() -> None:
+    op.drop_index("ix_code_shapes_snippet", table_name="code_shapes")
+    op.drop_index("ix_code_shapes_project_status", table_name="code_shapes")
+    op.drop_table("code_shapes")

frontend/src/views/ProjectView.vue

@@ -424,15 +424,17 @@ async function loadNotes() {
 
 interface CoverageGap {
   dir: string;
-  uncovered: number;
+  unclassified: number;
   total: number;
 }
 interface Coverage {
   total: number;
-  recorded: number;
+  accounted: number;
+  unclassified: number;
+  counts: Record<string, number>;
   estimate: boolean;
   computed_at: string;
-  repos: { repo: string; ref: string; total: number; recorded: number }[];
+  repos: { repo: string; ref: string; total: number; accounted: number }[];
   largest_gaps: CoverageGap[];
 }
 
@@ -712,32 +714,42 @@ async function confirmDelete() {
         </div>
         <template v-if="coverage">
           <div class="coverage-numbers">
-            <span class="coverage-count">{{ coverage.recorded }}/{{ coverage.total }}</span>
+            <span class="coverage-count">{{ coverage.accounted }}/{{ coverage.total }}</span>
-            <span class="coverage-label">shapes recorded</span>
+            <span class="coverage-label">shapes accounted for</span>
           </div>
           <div
             class="coverage-bar"
             role="progressbar"
-            :aria-valuenow="coverage.recorded"
+            :aria-valuenow="coverage.accounted"
             :aria-valuemin="0"
             :aria-valuemax="coverage.total"
-            aria-label="Shapes with a recorded snippet"
+            aria-label="Shapes classified against canon"
           >
             <div
               class="coverage-bar-fill"
-              :style="{ width: (coverage.total ? (coverage.recorded / coverage.total) * 100 : 0) + '%' }"
+              :style="{ width: (coverage.total ? (coverage.accounted / coverage.total) * 100 : 0) + '%' }"
             ></div>
           </div>
+          <div v-if="coverage.counts" class="coverage-gaps">
+            <span
+              v-for="k in ['canonical', 'instance', 'variant', 'exempt']"
+              :key="k"
+            >
+              <span v-if="coverage.counts[k]" class="coverage-gap-chip">
+                {{ coverage.counts[k] }} {{ k }}
+              </span>
+            </span>
+          </div>
           <div v-if="coverage.largest_gaps?.length" class="coverage-gaps">
-            <span class="coverage-gaps-label">Largest gaps:</span>
+            <span class="coverage-gaps-label">Most unclassified:</span>
             <span v-for="gap in coverage.largest_gaps" :key="gap.dir" class="coverage-gap-chip">
-              {{ gap.dir }} <span class="coverage-gap-count">{{ gap.uncovered }}</span>
+              {{ gap.dir }} <span class="coverage-gap-count">{{ gap.unclassified }}</span>
             </span>
           </div>
         </template>
         <p v-else class="coverage-empty">
-          Not measured yet — Refresh compares the bound repo's definitions
+          Not measured yet — Refresh reads the bound repo's definitions into
-          against recorded snippets.
+          the shape ledger and reports how many are classified against canon.
         </p>
         <p v-if="coverageError" class="coverage-error">{{ coverageError }}</p>
         <!-- Forge pin (#2778): owner-only, because the eligible set is the

plugin/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "scribe",
   "description": "Scribe system-of-record for Claude Code: MCP tools over your notes/tasks/projects/rules, a session-start push channel that surfaces your always-on rules + active-project context, process-skills (writing-plans, systematic-debugging, verification, brainstorming, reusing-code), and your saved Scribe Processes auto-surfaced as skills (/scribe:sync). Replaces superpowers + file-memory with one app-backed plugin.",
-  "version": "0.1.31",
+  "version": "0.1.32",
   "author": { "name": "Bryan Van Deusen" },
   "mcpServers": {
     "scribe": {

plugin/skills/reusing-code/SKILL.md

@@ -107,6 +107,21 @@ The result is a single entry that shows every place the thing is used — which
 exactly the signal that it was worth consolidating. This is the cure the create
 gate only hints at when it blocks a near-duplicate.
 
+## Consumer maps are rows, never prose
+
+Every enumerated relationship between code and canon belongs in the shape
+ledger, not in a sentence. When you establish that call sites route through a
+canonical helper — during an audit, a verify pass, or a consolidation —
+record each consuming definition with
+`classify_shapes(project_id, [{path, symbol, status: "instance", snippet_id}])`.
+A deliberate departure is a `"variant"` (reason required — the why IS the
+record); a judged one-off is `"exempt"` (reason required). Prose in a
+verification detail cannot be sorted, queried, or diffed; rows are what make
+"what uses this?" answerable forever. `list_shapes(project_id,
+status="unclassified")` is the standing todo — and N same-shaped occurrences
+matching no canon means derive one first (consolidate, `create_snippet`,
+then classify the rest against it), never N loose classifications.
+
 ## Why this pays off
 
 A one-off written a second time is the cost this avoids — and at project

src/scribe/mcp/server.py

@@ -49,15 +49,15 @@ Hierarchy: Project -> Milestone -> Task/Note. The map, by purpose:
 - WHERE work happens: Systems. Tag records with system_ids as you write;
   create_system when the area is unmodelled.
 - HOW to work: rules are pull-only and binding — call list_always_on_rules()
-  yourself at session start; a push that also delivered them was an
+  yourself at session start.
-  optimisation, not the bridge.
 - UI: the project's design system is binding — resolve_design_system /
   get_design_system_stylesheet before hand-writing a value.
 - REUSE: search snippets before writing a helper; record what you build with
-  create_snippet. Saved procedures are Processes (follow verbatim). Deletes
+  create_snippet; classify shapes against canon (classify_shapes) — a
-  are trash-recoverable.
+  consumer map is rows, never prose. Saved procedures are Processes (follow
+  verbatim). Deletes are trash-recoverable.
 
-A task is a note with status; *_note tools for notes, *_task for tasks.
+A task is a note with status (*_note vs *_task tools).
 Creates are duplicate-gated: a near-match BLOCKS and returns the existing
 id — update it, don't force. shared:true records are another user's — a
 suggestion, not the operator's settled practice.
@@ -110,6 +110,9 @@ _READ_ONLY_TOOLS = frozenset({
     # Which repos map to which project. Read-only by nature; bind_repo /
     # unbind_repo are the writes.
     "list_repo_bindings",
+    # The shape ledger's todo query (#2789). Reads only — classify_shapes is
+    # the write, and it is deliberately NOT here.
+    "list_shapes",
 })
 
 # Read-SHAPED tools that must NOT be reachable with a read key — a getter that

src/scribe/mcp/tools/__init__.py

@@ -5,8 +5,8 @@ to a FastMCP instance. `register_all(mcp)` is the single entry point called
 from `mcp.server.build_mcp_server`.
 """
 from scribe.mcp.tools import (
-    design_systems, milestones, notes, processes, projects, recent, repos, rulebooks, search, snippets,
+    design_systems, milestones, notes, processes, projects, recent, repos, rulebooks, search, shapes,
-    systems, tags, tasks, trash,
+    snippets, systems, tags, tasks, trash,
 )
 
 
@@ -24,5 +24,6 @@ def register_all(mcp) -> None:
     repos.register(mcp)
     processes.register(mcp)
     snippets.register(mcp)
+    shapes.register(mcp)
     rulebooks.register(mcp)
     trash.register(mcp)

src/scribe/mcp/tools/projects.py

@@ -61,11 +61,14 @@ async def enter_project(project_id: int) -> dict:
     open_tasks, recent_notes, design_system, systems, pattern_coverage —
     plus systems_bootstrap, present only when it applies (see below).
 
-    `pattern_coverage` (usually null) is a one-line estimate of how much of
+    `pattern_coverage` (usually null) is the shape-accounting line — how many
-    the bound repo's code has recorded snippets — e.g. "pattern-library
+    of the bound repo's extracted shapes carry a classification against canon
-    coverage: 34/210 shapes recorded (estimate); largest gaps: internal/api".
+    (note 2786) — e.g. "shape accounting: 3100/4573 shapes accounted for —
-    When present, treat the gaps as a standing invitation: as you touch code
+    12 canonical · 2900 instance (estimate, computed 2026-08-19); 1473
-    in those areas, record the shapes you find with create_snippet.
+    unclassified, largest: internal/api". Unclassified IS the todo: as you
+    touch code in those areas, classify the shapes you can (instances of
+    recorded canon, deliberate variants, one-off exemptions) and record the
+    canon that's missing with create_snippet.
 
     `systems` is the project's vocabulary of named subsystems/areas. It is
     returned here so you can TAG as you write: when creating or meaningfully
@@ -75,10 +78,10 @@ async def enter_project(project_id: int) -> dict:
     a subsystem's accumulated records with list_system_records.
 
     `systems_bootstrap` appears ONLY when the project has many records and no
-    Systems at all — act on it before starting other work: propose a starter
+    Systems at all — act on it before starting other work: create_system a
-    vocabulary from the areas the project's records name, confirm it with the
+    starter vocabulary from the areas the project's records name, directly
-    operator, and create_system the confirmed set. It stops appearing the
+    and without asking permission, preferring the standard names the ask
-    moment the first System exists.
+    lists. It stops appearing the moment the first System exists.
 
     `design_system` is null unless the project points at one. When present it
     carries the chain-merged guidance (the house style AND this project's
@@ -126,7 +129,7 @@ async def enter_project(project_id: int) -> dict:
 
     # The arrival-moment half of the bootstrap ask (#2683): session start is
     # when the agent has just read the project map and is not yet deep in a
-    # task — the one moment "propose a starter vocabulary" is cheap. The
+    # task — the one moment minting a starter vocabulary is cheap. The
     # write-moment half rides untagged-record responses (attach_systems);
     # both retire the instant the first System exists.
     systems_bootstrap = None

src/scribe/mcp/tools/shapes.py

@@ -0,0 +1,95 @@
+"""Shape-ledger MCP tools — the classification write/read surface (#2789).
+
+The accounting model (note 2786): the snippet library records CANON (small);
+the ledger accounts for EVERY extracted shape (total). These tools are how
+agents move shapes out of `unclassified` — the todo state — and how they read
+what still needs judgment. The ledger rows themselves are fed by the coverage
+refresh; these tools only ever judge what the sync has seen.
+"""
+from __future__ import annotations
+
+from scribe.mcp._context import current_user_id
+from scribe.services import shape_ledger as shape_ledger_svc
+
+
+async def classify_shapes(
+    project_id: int, classifications: list[dict], via: str = "agent"
+) -> dict:
+    """Record judgments for a project's code shapes — in batch, as rows.
+
+    EVERY shape in a bound repo should end up classified (note 2786):
+      - `instance` of snippet N — it conforms to recorded canon (family-level
+        canon in another project counts; that fully accounts for the shape).
+      - `variant` of snippet N — a deliberate, named departure. `reason`
+        (the why) is REQUIRED; it is the record.
+      - `exempt` — judged genuinely one-off. `reason` REQUIRED.
+      - `canonical` of snippet N — this row IS the snippet's reference
+        (rarely set by hand; the coverage sync stamps these mechanically).
+      - `unclassified` — withdraw a judgment; the shape rejoins the todo.
+
+    Consumer maps belong HERE, not in prose: when an audit enumerates call
+    sites of a canonical helper, each call site's defining shape is an
+    `instance` row — a sentence in a verification detail cannot be sorted,
+    queried, or diffed.
+
+    Args:
+        project_id: The project whose ledger is being judged.
+        classifications: Objects of {path, symbol, status, kind?, snippet_id?,
+            reason?}. path+symbol name the shape exactly as list_shapes shows
+            it; kind ("sym"/"css") narrows when one file defines both.
+            snippet_id is required for canonical/instance/variant; reason is
+            required for variant/exempt.
+        via: Who is judging — "agent" (default), "audit" (a sweep), or
+            "import" (carrying maps recorded elsewhere).
+
+    All-or-nothing: a structural error, a missing snippet target, or no write
+    access applies NOTHING. Returns {"classified": N, "unmatched": [...]} —
+    unmatched names shapes no live ledger row matches (the tree may have
+    moved since you listed; re-run the project's coverage refresh to re-sync).
+    """
+    uid = current_user_id()
+    return await shape_ledger_svc.classify_shapes(
+        uid, project_id, classifications, via=via
+    )
+
+
+async def list_shapes(
+    project_id: int,
+    status: str = "",
+    path: str = "",
+    snippet_id: int = 0,
+    include_vanished: bool = False,
+    limit: int = 100,
+    offset: int = 0,
+) -> dict:
+    """Read a project's shape ledger — `status="unclassified"` IS the todo.
+
+    Every extracted definition in the project's bound repos has a row here
+    (fed by the coverage refresh). Filters compose:
+
+    Args:
+        status: canonical | instance | variant | exempt | unclassified.
+        path: exact file, or a directory — matches everything beneath it
+            (the coverage line's "largest" dirs go straight in here).
+        snippet_id: rows classified against this snippet — a consumer map.
+        include_vanished: include shapes no longer in the tree (history).
+        limit/offset: page through big ledgers (limit caps at 500).
+
+    Returns {"shapes": [...], "total": N} — total counts every match, not
+    just this page. Classify what you can judge with classify_shapes; a
+    repeating shape with NO recorded canon is a derive-one-first moment
+    (consolidate onto a reference, create_snippet it, then classify the
+    rest against it), never N loose classifications.
+    """
+    uid = current_user_id()
+    rows, total = await shape_ledger_svc.list_project_shapes(
+        uid, project_id,
+        status=status, path=path, snippet_id=snippet_id,
+        include_vanished=include_vanished, limit=limit, offset=offset,
+    )
+    return {"shapes": [r.to_dict() for r in rows], "total": total}
+
+
+def register(mcp) -> None:
+    for fn in (classify_shapes, list_shapes):
+        mcp.tool(name=fn.__name__)(fn)

src/scribe/mcp/tools/snippets.py

@@ -207,6 +207,12 @@ async def get_snippet(snippet_id: int) -> dict:
     the source moved on — trust the location over the cached body and
     consider verify_snippet after you look.
 
+    When the shape ledger has judgments against this snippet, the response
+    carries `instances` (shapes classified as conforming to it — the
+    structured consumer map) and/or `variants` (named departures, each with
+    its why). Consult them before changing the snippet's contract: they are
+    the call sites your change lands on (classify_shapes maintains them).
+
     If the record belongs to someone else it carries `shared: true` with the
     `owner` and your `permission`. Read that as ONE PERSON'S SUGGESTION, not as
     established practice here: judge it on its merits, say whose it is when you
@@ -229,6 +235,16 @@ async def get_snippet(snippet_id: int) -> dict:
     await systems_tools.attach_systems(
         uid, note.user_id, data, note.id, note.project_id
     )
+    # The structured consumer map (#2789): ledger rows judged against this
+    # snippet. Attached only when non-empty (#2483) — and never for projects
+    # the caller can't read.
+    from scribe.services import shape_ledger as shape_ledger_svc
+
+    consumers = await shape_ledger_svc.snippet_consumers(uid, int(note.id))
+    if consumers["instances"]:
+        data["instances"] = consumers["instances"]
+    if consumers["variants"]:
+        data["variants"] = consumers["variants"]
     return data
 
 
@@ -336,6 +352,13 @@ async def verify_snippet(
     "moved to services/knowledge.py"). It's what makes the record fixable later
     by someone who wasn't here, so write it for them, not as a status echo.
 
+    CONSUMERS you enumerate while checking ("all N call sites still route
+    through it") are ledger rows, not detail prose: classify each consuming
+    definition as an `instance` of this snippet with classify_shapes — prose
+    cannot be sorted, queried, or diffed (note 2786's lesson), and the rows
+    are what make "what uses this?" answerable forever. `detail` keeps the
+    WHY and what changed, nothing that belongs in a row.
+
     A verdict expires automatically if the snippet is edited afterwards: it is
     stamped with a hash of the code it was checked against, so it can never go
     on vouching for code nobody checked. Re-verify after fixing a record.

src/scribe/mcp/tools/systems.py

@@ -23,6 +23,18 @@ from scribe.services import systems as systems_svc
 _BOOTSTRAP_MIN_RECORDS = 20
 _BOOTSTRAP_TITLES = 6
 
+# The standard vocabulary (#2798): area names that recur across software
+# projects, offered so "CI & Release" means the same thing in every project
+# on the instance. Consistency comes from the shared names — NOT from asking
+# the operator to approve each System; agents mint directly. Generic by
+# design (rule #115): archetypes any codebase could have, never one
+# install's subsystems. Mint freely beyond the list; the duplicate gate
+# guards sprawl.
+_STANDARD_SYSTEMS = (
+    "CI & Release", "Auth & Access", "Data Model & Storage", "API Surface",
+    "UI & Design", "Import & Export", "Background Jobs", "Observability",
+)
+
 
 async def bootstrap_systems_ask(user_id: int, project_id: int) -> str | None:
     """The escalated vocabulary-bootstrap ask for a mature zero-Systems project.
@@ -33,8 +45,10 @@ async def bootstrap_systems_ask(user_id: int, project_id: int) -> str | None:
     (#2683). What separates the nudges that convert from the prose that
     doesn't (the duplicate gate, the prior-art "already defined in 2 files")
     is the project's OWN evidence in the ask — so this one carries the record
-    count and the recent titles, and asks for a concrete deliverable: propose
+    count and the recent titles, and asks for a concrete deliverable: create
-    a starter set, confirm, create.
+    a starter set directly, preferring the standard cross-project names.
+    Deliberately NOT an approval flow (#2798): the operator is not a
+    permission gate for vocabulary; the standard names carry consistency.
 
     Self-retiring by construction: callers only reach for it while the
     project has zero Systems, so the first create_system ends it everywhere.
@@ -56,12 +70,15 @@ async def bootstrap_systems_ask(user_id: int, project_id: int) -> str | None:
     return (
         f"This project has {total} records and NO Systems modelled — none of "
         "them can be tagged to an area, so recurring problem-spots stay "
-        "invisible. Bootstrap the vocabulary now, in this session: from the "
+        "invisible. Bootstrap the vocabulary now, in this session, without "
-        f"areas the records themselves name (recent: {titles}), propose 3-6 "
+        "asking permission — creating Systems is your call, not an approval "
-        "Systems to the operator, create_system each confirmed one with a "
+        f"flow. From the areas the records themselves name (recent: {titles}), "
-        "one-paragraph charter, then tag this record (system_ids=[...]). "
+        "create_system 3-6 Systems, each with a one-paragraph charter, then "
-        "This ask repeats until the first System exists; answering it once "
+        "tag this record (system_ids=[...]). Where an area fits a standard "
-        "retires it for every future record."
+        f"name, use it verbatim so it means the same thing in every project: "
+        f"{', '.join(_STANDARD_SYSTEMS)}. Mint freely beyond that list — the "
+        "duplicate gate guards sprawl. This ask repeats until the first "
+        "System exists; answering it once retires it for every future record."
     )
 
 
@@ -148,11 +165,16 @@ async def create_system(
 
     Create one the moment two records would share an area that has no System
     yet — the same two-or-more test snippets use. Don't wait to be asked to
-    name an area that plainly exists in the code; an unmodelled area means
+    name an area that plainly exists in the code, and don't route the
-    every record about it stays untaggable. An audit or sweep that walks the
+    creation through operator approval — minting vocabulary is the agent's
-    codebase is a DISCOVERY moment: mint the Systems it names as it names
+    call (#2798); an unmodelled area means every record about it stays
-    them — the duplicate gate below, plus reviewing the existing list, is what
+    untaggable. An audit or sweep that walks the codebase is a DISCOVERY
-    guards against sprawl, not holding back. Give each one a one-paragraph
+    moment: mint the Systems it names as it names them — the duplicate gate
+    below, plus reviewing the existing list, is what guards against sprawl,
+    not holding back. Prefer the standard cross-project names where the area
+    fits one (CI & Release, Auth & Access, Data Model & Storage, API Surface,
+    UI & Design, Import & Export, Background Jobs, Observability) so the same
+    word means the same thing in every project. Give each one a one-paragraph
     charter, not just a label: the description is what tells a later session
     whether a record belongs here.
 

src/scribe/models/__init__.py

@@ -44,5 +44,6 @@ from scribe.models.rulebook import (  # noqa: E402, F401
 )
 from scribe.models.repo_binding import RepoBinding  # noqa: E402, F401
 from scribe.models.forge_connection import ForgeConnection  # noqa: E402, F401
+from scribe.models.code_shape import CodeShape  # noqa: E402, F401
 from scribe.models.system import System, RecordSystem  # noqa: E402, F401
 from scribe.models.design_system import DesignSystem, DesignToken  # noqa: E402, F401

src/scribe/models/code_shape.py

@@ -0,0 +1,98 @@
+from datetime import datetime
+
+from sqlalchemy import (
+    BigInteger,
+    DateTime,
+    ForeignKey,
+    Index,
+    Integer,
+    Text,
+    UniqueConstraint,
+)
+from sqlalchemy.orm import Mapped, mapped_column
+
+from scribe.models import Base
+from scribe.models.base import TimestampMixin
+
+# The classification vocabulary (note 2786). `unclassified` is the default and
+# THE todo state; every other status is a judgment, stamped with who made it.
+SHAPE_STATUSES = ("canonical", "instance", "variant", "exempt", "unclassified")
+SHAPE_CLASSIFIERS = ("agent", "audit", "hook", "mechanical", "import")
+
+
+class CodeShape(Base, TimestampMixin):
+    """One extracted code shape and its classification against canon (#2787).
+
+    The accounting half of the pattern system (governing note 2786): the
+    snippet library records CANON (small); this ledger accounts for EVERY
+    shape the coverage extractor finds in a bound repo (total). A row's
+    status says how the shape relates to canon — it IS a snippet's reference
+    (`canonical`), conforms to one (`instance` — snippet_id may point at
+    another project's snippet, so family canon counts), departs deliberately
+    (`variant`, with the why in `reason`), was judged one-off (`exempt`,
+    a recorded judgment rather than silence), or awaits judgment
+    (`unclassified` — the todo).
+
+    Identity is (project, repo_key, path, symbol, kind) — kind is part of it
+    because one file can define `.foo` (css) and `foo` (sym) as distinct
+    shapes. A rename therefore reads as vanish + new row: accepted for v1,
+    because chasing renames needs content identity the extractor doesn't
+    have. `vanished_at` keeps the history instead of deleting it.
+
+    snippet_id is SET NULL on snippet deletion: the classification's target
+    is gone but the judgment happened; the sync pass (step 2) re-files such
+    rows as unclassified so they rejoin the todo instead of dangling.
+    """
+
+    __tablename__ = "code_shapes"
+    __table_args__ = (
+        UniqueConstraint(
+            "project_id", "repo_key", "path", "symbol", "kind",
+            name="uq_code_shapes_identity",
+        ),
+        Index("ix_code_shapes_project_status", "project_id", "status"),
+        Index("ix_code_shapes_snippet", "snippet_id"),
+    )
+
+    id: Mapped[int] = mapped_column(primary_key=True)
+    project_id: Mapped[int] = mapped_column(
+        Integer, ForeignKey("projects.id", ondelete="CASCADE"), nullable=False
+    )
+    repo_key: Mapped[str] = mapped_column(Text, nullable=False)
+    path: Mapped[str] = mapped_column(Text, nullable=False)
+    symbol: Mapped[str] = mapped_column(Text, nullable=False)
+    kind: Mapped[str] = mapped_column(Text, nullable=False)  # "css" | "sym"
+    status: Mapped[str] = mapped_column(Text, nullable=False, default="unclassified")
+    snippet_id: Mapped[int | None] = mapped_column(
+        BigInteger, ForeignKey("notes.id", ondelete="SET NULL"), nullable=True
+    )
+    reason: Mapped[str | None] = mapped_column(Text, nullable=True)
+    classified_by: Mapped[str | None] = mapped_column(Text, nullable=True)
+    classified_at: Mapped[datetime | None] = mapped_column(
+        DateTime(timezone=True), nullable=True
+    )
+    first_seen_commit: Mapped[str] = mapped_column(Text, default="")
+    last_seen_commit: Mapped[str] = mapped_column(Text, default="")
+    vanished_at: Mapped[datetime | None] = mapped_column(
+        DateTime(timezone=True), nullable=True
+    )
+
+    def to_dict(self) -> dict:
+        return {
+            "id": self.id,
+            "project_id": self.project_id,
+            "repo_key": self.repo_key,
+            "path": self.path,
+            "symbol": self.symbol,
+            "kind": self.kind,
+            "status": self.status,
+            "snippet_id": self.snippet_id,
+            "reason": self.reason,
+            "classified_by": self.classified_by,
+            "classified_at": self.classified_at.isoformat() if self.classified_at else None,
+            "first_seen_commit": self.first_seen_commit,
+            "last_seen_commit": self.last_seen_commit,
+            "vanished_at": self.vanished_at.isoformat() if self.vanished_at else None,
+            "created_at": self.created_at.isoformat(),
+            "updated_at": self.updated_at.isoformat(),
+        }

src/scribe/routes/projects.py

@@ -123,7 +123,7 @@ async def delete_project_route(project_id: int):
 @projects_bp.route("/<int:project_id>/coverage", methods=["GET"])
 @login_required
 async def get_coverage_route(project_id: int):
-    """The cached pattern-library coverage summary — never computes.
+    """The cached shape-accounting summary — never computes.
 
     `configured` tells the card whether offering a Refresh button makes
     sense; `coverage` is null until something has computed it (a webhook

src/scribe/services/backup.py

@@ -11,6 +11,7 @@ from scribe.models.note_supersession import NoteSupersession
 from scribe.models.note_version import NoteVersion
 from scribe.models.design_system import DesignSystem, DesignToken
 from scribe.models.note_usage import NoteUsageEvent
+from scribe.models.code_shape import CodeShape
 from scribe.models.project import Project
 from scribe.models.repo_binding import RepoBinding
 from scribe.models.rulebook import (
@@ -36,8 +37,11 @@ logger = logging.getLogger(__name__)
 # v6 (2026-08) added note_supersessions — and the guard did stop the seventh:
 # the table shipped without a backup section and the coverage test failed the
 # build, which is the whole reason that list was written.
+# v7 (2026-08) added code_shapes — the shape ledger (#2787). Classifications
+# are judgment data worth carrying; a restore keeps a judgment only when its
+# snippet target survives the id re-mapping, else the row rejoins the todo.
 # Bump when the serialized schema changes.
-BACKUP_VERSION = 6
+BACKUP_VERSION = 7
 
 # Every table this backup carries, by its REAL name. Paired with _NOT_INCLUDED
 # below, these two lists must together account for the entire schema — which is
@@ -55,6 +59,8 @@ _BACKED_UP = [
     # v5 (2026-08): the five-year gap this list was written to stop.
     "systems", "record_systems", "design_systems", "design_tokens",
     "note_usage_events", "repo_bindings", "note_supersessions",
+    # v7 (2026-08): the shape ledger (#2787).
+    "code_shapes",
 ]
 
 # Tables intentionally NOT in the backup, surfaced in the payload so the gap is
@@ -166,6 +172,10 @@ def _usage_event_rows(rows) -> list[dict]:
     ]
 
 
+def _code_shape_rows(rows) -> list[dict]:
+    return [r.to_dict() for r in rows]
+
+
 def _repo_binding_rows(rows) -> list[dict]:
     return [
         {"user_id": r.user_id, "project_id": r.project_id, "repo_key": r.repo_key}
@@ -205,6 +215,7 @@ async def export_full_backup() -> dict:
         design_tokens = (await session.execute(select(DesignToken))).scalars().all()
         usage_events = (await session.execute(select(NoteUsageEvent))).scalars().all()
         repo_bindings = (await session.execute(select(RepoBinding))).scalars().all()
+        code_shapes = (await session.execute(select(CodeShape))).scalars().all()
         rulebooks = (await session.execute(select(Rulebook))).scalars().all()
         topics = (await session.execute(select(RulebookTopic))).scalars().all()
         rules = (await session.execute(select(Rule))).scalars().all()
@@ -379,6 +390,7 @@ async def export_full_backup() -> dict:
         "note_usage_events": _usage_event_rows(usage_events),
         "repo_bindings": _repo_binding_rows(repo_bindings),
         "note_supersessions": _note_supersession_rows(supersessions),
+        "code_shapes": _code_shape_rows(code_shapes),
     }
 
 
@@ -446,6 +458,11 @@ async def export_user_backup(user_id: int) -> dict:
         repo_bindings = (await session.execute(
             select(RepoBinding).where(RepoBinding.user_id == user_id)
         )).scalars().all()
+        # The ledger has no user_id of its own — rows belong to the project
+        # they account for, so a user's export carries their projects' rows.
+        code_shapes = (await session.execute(
+            select(CodeShape).where(CodeShape.project_id.in_(project_ids))
+        )).scalars().all() if project_ids else []
         rulebooks = (await session.execute(
             select(Rulebook).where(Rulebook.owner_user_id == user_id)
         )).scalars().all()
@@ -634,6 +651,7 @@ async def export_user_backup(user_id: int) -> dict:
         "note_usage_events": _usage_event_rows(usage_events),
         "repo_bindings": _repo_binding_rows(repo_bindings),
         "note_supersessions": _note_supersession_rows(supersessions),
+        "code_shapes": _code_shape_rows(code_shapes),
     }
 
 
@@ -737,7 +755,7 @@ async def _restore_v2(data: dict) -> dict:
         "topic_suppressions": 0,
         "systems": 0, "record_systems": 0, "design_systems": 0,
         "design_tokens": 0, "note_usage_events": 0, "repo_bindings": 0,
-        "note_supersessions": 0,
+        "note_supersessions": 0, "code_shapes": 0,
     }
 
     async with async_session() as session:
@@ -1114,6 +1132,42 @@ async def _restore_v2(data: dict) -> dict:
             ))
             stats["repo_bindings"] += 1
 
+        # 21. Code shapes (v7, #2787) — the ledger's classifications are
+        # judgments worth carrying. A judgment whose snippet target didn't
+        # survive the re-mapping (canonical/instance/variant with a gone
+        # snippet) is downgraded to unclassified so it rejoins the todo
+        # honestly instead of dangling; exempt needs no target and keeps.
+        for cs_data in data.get("code_shapes", []):
+            mapped_pid = project_id_map.get(cs_data.get("project_id", 0))
+            if mapped_pid is None:
+                continue
+            status = cs_data.get("status", "unclassified")
+            mapped_sid = note_id_map.get(cs_data.get("snippet_id") or 0)
+            classified_by = cs_data.get("classified_by")
+            classified_at = cs_data.get("classified_at")
+            if status in ("canonical", "instance", "variant") and mapped_sid is None:
+                status = "unclassified"
+                classified_by = None
+                classified_at = None
+            session.add(CodeShape(
+                project_id=mapped_pid,
+                repo_key=cs_data.get("repo_key", ""),
+                path=cs_data.get("path", ""),
+                symbol=cs_data.get("symbol", ""),
+                kind=cs_data.get("kind", "sym"),
+                status=status,
+                snippet_id=mapped_sid,
+                reason=cs_data.get("reason"),
+                classified_by=classified_by,
+                classified_at=_dt(classified_at) if classified_at else None,
+                first_seen_commit=cs_data.get("first_seen_commit", ""),
+                last_seen_commit=cs_data.get("last_seen_commit", ""),
+                vanished_at=_dt(cs_data["vanished_at"]) if cs_data.get("vanished_at") else None,
+                created_at=_dt(cs_data.get("created_at")),
+                updated_at=_dt(cs_data.get("updated_at")),
+            ))
+            stats["code_shapes"] += 1
+
         await session.commit()
 
     logger.info("Restored v2/v3 backup: %s", stats)

src/scribe/services/coverage.py

@@ -41,7 +41,10 @@ logger = logging.getLogger(__name__)
 
 # Cache key in the settings KV, on the project OWNER's user_id — the same
 # channel the scheduler's last-run summary uses for machine-written state.
-_CACHE_KEY_PREFIX = "pattern_coverage_"
+# v2 suffix with #2788: the payload shape inverted (accounted/unclassified);
+# pre-ledger blobs under the old key simply stop being found, so the card
+# honestly reads "not measured yet" until the first ledger-era refresh.
+_CACHE_KEY_PREFIX = "pattern_coverage_v2_"
 
 # Files whose content can't hold definitions — the hook's skip list, verbatim,
 # plus sourcemaps (which are JSON in a trenchcoat).
@@ -157,77 +160,42 @@ def shapes_from_archive(blob: bytes) -> list[tuple[str, str, str]]:
 # --- matching shapes against recorded locations ------------------------------
 
 
-def _norm_symbol(kind_or_symbol: str) -> str:
+# The covering predicate (location_covers) lives in services/shape_ledger.py
-    # CSS shapes and recorded CSS symbols may or may not carry the leading
+# since #2788 — the ledger's canonical marking and this module's readout are
-    # dot; compare without it so ".btn-primary" and "btn-primary" agree.
+# two consumers of ONE doctrine, and the ledger is its home.
-    return kind_or_symbol.lstrip(".").strip()
-
-
-def _location_covers(loc_path: str, loc_symbol: str, path: str, name: str) -> bool:
-    if _norm_symbol(loc_symbol) != _norm_symbol(name):
-        return False
-    if not loc_path:
-        # Symbol-only record: the symbol match is all the claim there is.
-        return True
-    # The drift check's location semantics, not a second copy of them: exact
-    # file, or the recorded path is a directory the file lives under.
-    from scribe.services.snippets import _path_touches
-
-    return _path_touches(loc_path, path)
-
-
-def match_shapes(
-    shapes: list[tuple[str, str, str]],
-    recorded: list[tuple[str, str]],
-) -> list[tuple[str, str, str, bool]]:
-    """Each shape with whether some recorded (path, symbol) location covers it.
-
-    Symbol-less recorded locations never cover a shape — a whole-file record
-    makes no claim about any particular definition inside it. The recorded
-    repo NAME is deliberately not consulted: it is free-form ("Scribe") and
-    the project binding already did the scoping; on a project binding several
-    repos this can over-credit a same-named symbol, which the estimate label
-    owns.
-    """
-    usable = [(p, s) for p, s in recorded if (s or "").strip()]
-    return [
-        (
-            path,
-            kind,
-            name,
-            any(_location_covers(lp, ls, path, name) for lp, ls in usable),
-        )
-        for path, kind, name in shapes
-    ]
 
 
 def largest_gaps(
-    matched: list[tuple[str, str, str, bool]], *, top: int = 3
+    accounted: list[tuple[str, str, str, bool]], *, top: int = 3
 ) -> list[dict]:
-    """The directories with the most uncovered shapes — where a backlog
+    """The directories with the most unclassified shapes — where a
-    session should start, named the way the repo names them."""
+    classification session should start, named the way the repo names them."""
     by_dir: dict[str, dict[str, int]] = {}
-    for path, _kind, _name, covered in matched:
+    for path, _kind, _name, is_accounted in accounted:
         d = posixpath.dirname(path) or "(root)"
-        row = by_dir.setdefault(d, {"total": 0, "uncovered": 0})
+        row = by_dir.setdefault(d, {"total": 0, "unclassified": 0})
         row["total"] += 1
-        if not covered:
+        if not is_accounted:
-            row["uncovered"] += 1
+            row["unclassified"] += 1
     ranked = sorted(
-        by_dir.items(), key=lambda kv: (-kv[1]["uncovered"], kv[0])
+        by_dir.items(), key=lambda kv: (-kv[1]["unclassified"], kv[0])
     )
     return [
-        {"dir": d, "uncovered": row["uncovered"], "total": row["total"]}
+        {"dir": d, "unclassified": row["unclassified"], "total": row["total"]}
         for d, row in ranked[:top]
-        if row["uncovered"]
+        if row["unclassified"]
     ]
 
 
 # --- compute, cache, surface -------------------------------------------------
 
 
-async def _recorded_locations(user_id: int, project_id: int) -> list[tuple[str, str]]:
+async def _recorded_locations(
-    """(path, symbol) for every location of every live snippet in a project."""
+    user_id: int, project_id: int
+) -> list[tuple[int, str, str]]:
+    """(snippet_note_id, path, symbol) for every location of every live
+    snippet in a project — the canonical-marking input (#2788): the id is what
+    lets a ledger row point back at the snippet it references."""
     from sqlalchemy import select
 
     from scribe.models import async_session
@@ -244,34 +212,46 @@ async def _recorded_locations(user_id: int, project_id: int) -> list[tuple[str,
             )
         )
         notes = list(rows.scalars().all())
-    out: list[tuple[str, str]] = []
+    out: list[tuple[int, str, str]] = []
     for note in notes:
         for loc in snippet_fields(note).get("locations") or []:
-            out.append((loc.get("path") or "", loc.get("symbol") or ""))
+            out.append(
+                (int(note.id), loc.get("path") or "", loc.get("symbol") or "")
+            )
     return out
 
 
 async def compute_coverage(
     user_id: int, project_id: int, *, selector: ForgeSelector | None = None
 ) -> dict | None:
-    """Measure a project's pattern-library coverage against its bound repos.
+    """Sync the shape ledger from the bound repos and read the accounting.
+
+    Since #2788 this is the ledger's sync point, not just a measurement:
+    every walk upserts the extracted shapes (new → unclassified, vanished →
+    stamped), re-stamps snippet reference locations as canonical, and then
+    reports the ACCOUNTING — how many shapes carry a classification at all —
+    rather than the old "has a snippet" fraction. Unclassified is the todo
+    (note 2786).
 
     None means "nothing to measure" — the owner's keyring serves none of the
     project's bound repos (#2778). That is the ordinary state for a
-    forge-less user and every caller treats it as silence, not failure.
+    forge-less user and every caller treats it as silence, not failure; the
-    Forge errors (unreachable, bad token) RAISE — the two callers are a
+    ledger is untouched in that case. Forge errors (unreachable, bad token)
-    refresh button and a background task, and both want to know.
+    RAISE — the two callers are a refresh button and a background task, and
+    both want to know.
 
     ``user_id`` is the project OWNER's id: the cache lives there, and the
     keyring resolved here must be the same one every other read uses.
     """
+    from scribe.services import shape_ledger
+    from scribe.services.forge import ForgeError
+
     if selector is None:
         selector = await get_forges(user_id, project_id)
     if not selector.configured:
         return None
 
-    repos: list[dict] = []
+    served: list[tuple[str, str]] = []
-    matched_all: list[tuple[str, str, str, bool]] = []
     recorded = await _recorded_locations(user_id, project_id)
     for key in await keys_for_project(user_id, project_id):
         hit = selector.resolve(key)
@@ -280,26 +260,54 @@ async def compute_coverage(
         forge, api_repo = hit
         ref = await forge.default_branch(api_repo)
         shapes = shapes_from_archive(await forge.archive(api_repo, ref))
-        matched = match_shapes(shapes, recorded)
+        # The head commit is provenance sugar on the ledger rows; failing to
-        matched_all.extend(matched)
+        # learn it must not fail the sync — the ref names the point well
-        repos.append({
+        # enough and the row timestamps carry the when.
-            "repo": key,
+        try:
-            "ref": ref,
+            marker = await forge.latest_commit(api_repo, "", ref) or ref
-            "total": len(matched),
+        except ForgeError:
-            "recorded": sum(1 for *_x, covered in matched if covered),
+            marker = ref
-        })
+        await shape_ledger.sync_repo_shapes(
-    if not repos:
+            project_id, key, shapes, seen_marker=marker
+        )
+        served.append((key, ref))
+    if not served:
         return None
 
+    await shape_ledger.mark_canonicals(project_id, recorded)
+
+    # Project-wide readout, deliberately wider than this walk: a second bound
+    # repo that was unreachable today still has live rows, and they count.
+    rows = await shape_ledger.live_rows(project_id)
+    counts = {"canonical": 0, "instance": 0, "variant": 0, "exempt": 0,
+              "unclassified": 0}
+    for row in rows:
+        counts[row.status] = counts.get(row.status, 0) + 1
+    by_repo: dict[str, dict[str, int]] = {}
+    for row in rows:
+        agg = by_repo.setdefault(row.repo_key, {"total": 0, "accounted": 0})
+        agg["total"] += 1
+        agg["accounted"] += row.status != "unclassified"
+
+    unclassified = counts.pop("unclassified")
     return {
-        "total": len(matched_all),
+        "total": len(rows),
-        "recorded": sum(1 for *_x, covered in matched_all if covered),
+        "accounted": len(rows) - unclassified,
+        "unclassified": unclassified,
+        "counts": counts,
         # Honesty flag, not decoration: every surface that shows the number
         # is expected to carry it through.
         "estimate": True,
         "computed_at": datetime.now(timezone.utc).isoformat(),
-        "repos": repos,
+        "repos": [
-        "largest_gaps": largest_gaps(matched_all),
+            {"repo": key, "ref": ref,
+             **by_repo.get(key, {"total": 0, "accounted": 0})}
+            for key, ref in served
+        ],
+        "largest_gaps": largest_gaps([
+            (r.path, r.kind, r.symbol, r.status != "unclassified")
+            for r in rows
+        ]),
     }
 
 
@@ -330,12 +338,23 @@ async def cached_coverage(user_id: int, project_id: int) -> dict | None:
 def coverage_line(coverage: dict) -> str:
     """The one-line evidence-carrying summary enter_project surfaces."""
     day = (coverage.get("computed_at") or "")[:10]
-    line = (
+    counts = coverage.get("counts") or {}
-        f"pattern-library coverage: {coverage.get('recorded', 0)}"
+    breakdown = " · ".join(
-        f"/{coverage.get('total', 0)} shapes recorded"
+        f"{counts[k]} {k}"
-        f" (estimate{', computed ' + day if day else ''})"
+        for k in ("canonical", "instance", "variant", "exempt")
+        if counts.get(k)
     )
+    line = (
+        f"shape accounting: {coverage.get('accounted', 0)}"
+        f"/{coverage.get('total', 0)} shapes accounted for"
+    )
+    if breakdown:
+        line += f" — {breakdown}"
+    line += f" (estimate{', computed ' + day if day else ''})"
+    unclassified = coverage.get("unclassified", 0)
+    if unclassified:
+        line += f"; {unclassified} unclassified"
         gaps = [g["dir"] for g in coverage.get("largest_gaps") or []]
         if gaps:
-        line += "; largest gaps: " + ", ".join(gaps)
+            line += ", largest: " + ", ".join(gaps)
     return line

src/scribe/services/shape_ledger.py

@@ -0,0 +1,410 @@
+"""The shape ledger's write side — sync and mechanical marking (#2788).
+
+The coverage walk (services/coverage.py) is the only feed that sees every
+shape, so it is the ledger's sync point: each refresh upserts one repo's
+extracted shapes — new shapes arrive `unclassified` (THE todo state, note
+2786), surviving shapes bump their last-seen marker, vanished shapes get
+stamped rather than deleted (history is the point). Classifications survive
+recompute by construction: the upsert never touches a judgment, with two
+deliberate exceptions —
+
+  - a judgment whose snippet target is gone (SET NULL on snippet deletion)
+    is re-filed as unclassified so it rejoins the todo instead of dangling;
+  - a MECHANICALLY-stamped canonical row whose snippet location no longer
+    covers it falls back to unclassified. Only mechanical stamps self-heal;
+    an agent's judgment is never unwound by machinery.
+
+`location_covers` is the one covering predicate — the same doctrine the
+recorded-location drift check uses — shared by the sync's canonical marking
+and by anything else that must decide whether a recorded location speaks for
+an extracted shape.
+"""
+from __future__ import annotations
+
+from datetime import datetime, timezone
+
+from sqlalchemy import select
+
+from scribe.models import async_session
+from scribe.models.code_shape import CodeShape
+
+# Statuses whose meaning requires a snippet target.
+_NEEDS_TARGET = ("canonical", "instance", "variant")
+
+
+def _norm_symbol(kind_or_symbol: str) -> str:
+    # CSS shapes and recorded CSS symbols may or may not carry the leading
+    # dot; compare without it so ".btn-primary" and "btn-primary" agree.
+    return kind_or_symbol.lstrip(".").strip()
+
+
+def location_covers(loc_path: str, loc_symbol: str, path: str, name: str) -> bool:
+    """Does a recorded (path, symbol) location speak for this shape?
+
+    Symbol-less locations never cover a shape — a whole-file record makes no
+    claim about any particular definition inside it. Path semantics are the
+    drift check's own: exact file, or the recorded path is a directory the
+    file lives under.
+    """
+    if not (loc_symbol or "").strip():
+        return False
+    if _norm_symbol(loc_symbol) != _norm_symbol(name):
+        return False
+    if not loc_path:
+        # Symbol-only record: the symbol match is all the claim there is.
+        return True
+    from scribe.services.snippets import _path_touches
+
+    return _path_touches(loc_path, path)
+
+
+async def sync_repo_shapes(
+    project_id: int,
+    repo_key: str,
+    shapes: list[tuple[str, str, str]],
+    *,
+    seen_marker: str,
+) -> None:
+    """Upsert one repo's extracted (path, kind, name) shapes into the ledger.
+
+    ``seen_marker`` is the commit the archive was read at when the forge can
+    say, else the ref name — provenance sugar; the row timestamps carry the
+    when.
+    """
+    now = datetime.now(timezone.utc)
+    async with async_session() as session:
+        rows = (
+            await session.execute(
+                select(CodeShape).where(
+                    CodeShape.project_id == project_id,
+                    CodeShape.repo_key == repo_key,
+                )
+            )
+        ).scalars().all()
+        by_key = {(r.path, r.symbol, r.kind): r for r in rows}
+        seen: set[tuple[str, str, str]] = set()
+        for path, kind, name in shapes:
+            key = (path, name, kind)
+            if key in seen:
+                continue
+            seen.add(key)
+            row = by_key.get(key)
+            if row is None:
+                session.add(CodeShape(
+                    project_id=project_id, repo_key=repo_key,
+                    path=path, symbol=name, kind=kind,
+                    first_seen_commit=seen_marker, last_seen_commit=seen_marker,
+                ))
+                continue
+            row.last_seen_commit = seen_marker
+            # A shape that vanished and came back is live again — the vanish
+            # stays visible in history via updated_at, not as a dead flag.
+            row.vanished_at = None
+            if row.status in _NEEDS_TARGET and row.snippet_id is None:
+                row.status = "unclassified"
+                row.classified_by = None
+                row.classified_at = None
+                row.reason = None
+        for key, row in by_key.items():
+            if key not in seen and row.vanished_at is None:
+                row.vanished_at = now
+        await session.commit()
+
+
+async def mark_canonicals(
+    project_id: int, recorded: list[tuple[int, str, str]]
+) -> None:
+    """Stamp snippet reference locations as `canonical` — the one mechanical
+    rule that is always safe (the judgment happened when the snippet was
+    minted; this row just makes it queryable).
+
+    ``recorded`` is (snippet_note_id, path, symbol) for every live snippet
+    location in the project. Touches only rows machinery owns: unclassified
+    rows gain the stamp; mechanically-stamped canonicals no longer covered
+    fall back to unclassified. Agent judgments are never overwritten.
+    """
+    usable = [(nid, p, s) for nid, p, s in recorded if (s or "").strip()]
+    now = datetime.now(timezone.utc)
+    async with async_session() as session:
+        rows = (
+            await session.execute(
+                select(CodeShape).where(
+                    CodeShape.project_id == project_id,
+                    CodeShape.vanished_at.is_(None),
+                )
+            )
+        ).scalars().all()
+        for row in rows:
+            covering = next(
+                (
+                    nid for nid, lp, ls in usable
+                    if location_covers(lp, ls, row.path, row.symbol)
+                ),
+                None,
+            )
+            if covering is not None and row.status == "unclassified":
+                row.status = "canonical"
+                row.snippet_id = covering
+                row.classified_by = "mechanical"
+                row.classified_at = now
+            elif (
+                covering is None
+                and row.status == "canonical"
+                and row.classified_by == "mechanical"
+            ):
+                row.status = "unclassified"
+                row.snippet_id = None
+                row.classified_by = None
+                row.classified_at = None
+        await session.commit()
+
+
+async def live_rows(project_id: int) -> list[CodeShape]:
+    """Every un-vanished ledger row for a project — the accounting readout's
+    input, across ALL its repos (a repo unreachable this refresh still counts;
+    accounting is project-wide)."""
+    async with async_session() as session:
+        return list(
+            (
+                await session.execute(
+                    select(CodeShape).where(
+                        CodeShape.project_id == project_id,
+                        CodeShape.vanished_at.is_(None),
+                    )
+                )
+            ).scalars().all()
+        )
+
+
+# --- classification (#2789): the judgment write path --------------------------
+
+# Statuses an explicit classification may set. All five: setting a row back to
+# `unclassified` is how a judgment is deliberately withdrawn.
+_SETTABLE = ("canonical", "instance", "variant", "exempt", "unclassified")
+
+# Who may appear as the classifier on this path. `hook` and `mechanical` are
+# server-internal feeds (steps 5-6) — a caller claiming them would launder a
+# judgment as machinery.
+_CALLER_VIAS = ("agent", "audit", "import")
+
+
+def validate_classifications(items: list[dict]) -> str | None:
+    """The structural error a classification batch would earn, or None.
+
+    Pure and checked BEFORE anything is touched: a batch either applies or
+    errors whole — the StrictArgs lesson (#2709), a caller must never learn
+    later that half a batch silently happened.
+    """
+    if not items:
+        return "classifications is empty — nothing to apply"
+    for i, item in enumerate(items):
+        if not isinstance(item, dict):
+            return f"classifications[{i}] is not an object"
+        path = (item.get("path") or "").strip()
+        symbol = (item.get("symbol") or "").strip()
+        if not path or not symbol:
+            return f"classifications[{i}] needs both path and symbol"
+        status = item.get("status") or ""
+        if status not in _SETTABLE:
+            return (
+                f"classifications[{i}] has unknown status {status!r} "
+                f"(one of: {', '.join(_SETTABLE)})"
+            )
+        snippet_id = item.get("snippet_id") or 0
+        if status in _NEEDS_TARGET and not snippet_id:
+            return (
+                f"classifications[{i}]: status {status!r} needs snippet_id — "
+                "the snippet this shape is (or departs from)"
+            )
+        if status in ("variant", "exempt") and not (item.get("reason") or "").strip():
+            return (
+                f"classifications[{i}]: status {status!r} needs a reason — "
+                "the WHY is the record (note 2786)"
+            )
+    return None
+
+
+async def classify_shapes(
+    user_id: int,
+    project_id: int,
+    classifications: list[dict],
+    *,
+    via: str = "agent",
+) -> dict:
+    """Apply a batch of judgments to a project's live ledger rows.
+
+    All-or-nothing on errors: the whole batch is validated (structure, write
+    access, every snippet target readable by the caller) before any row is
+    touched. Rows are matched by exact (path, symbol) — plus kind when the
+    item carries one — and a target no live row matches is reported in
+    ``unmatched``, not an error: the tree may simply have moved since the
+    caller listed. Idempotent by construction.
+    """
+    from scribe.services import access
+    from scribe.services import snippets as snippets_svc
+
+    if via not in _CALLER_VIAS:
+        raise ValueError(f"via must be one of: {', '.join(_CALLER_VIAS)}")
+    error = validate_classifications(classifications)
+    if error:
+        raise ValueError(error)
+    if not await access.can_write_project(user_id, project_id):
+        raise ValueError(f"project {project_id} not found or no write access")
+
+    # Snippet targets resolve through the caller's own read access — a
+    # family-canon snippet in another project counts (note 2786), a snippet
+    # the caller cannot read does not exist for them.
+    target_ids = {
+        int(item["snippet_id"])
+        for item in classifications
+        if item.get("status") in _NEEDS_TARGET
+    }
+    for sid in sorted(target_ids):
+        if await snippets_svc.get_snippet(user_id, sid) is None:
+            raise ValueError(f"snippet {sid} not found (or not readable)")
+
+    now = datetime.now(timezone.utc)
+    classified = 0
+    unmatched: list[dict] = []
+    async with async_session() as session:
+        rows = (
+            await session.execute(
+                select(CodeShape).where(
+                    CodeShape.project_id == project_id,
+                    CodeShape.vanished_at.is_(None),
+                )
+            )
+        ).scalars().all()
+        by_key: dict[tuple[str, str], list[CodeShape]] = {}
+        for row in rows:
+            by_key.setdefault((row.path, row.symbol), []).append(row)
+        for item in classifications:
+            matches = by_key.get(
+                ((item.get("path") or "").strip(), (item.get("symbol") or "").strip())
+            ) or []
+            kind = (item.get("kind") or "").strip()
+            if kind:
+                matches = [r for r in matches if r.kind == kind]
+            if not matches:
+                unmatched.append({
+                    "path": item.get("path"), "symbol": item.get("symbol"),
+                })
+                continue
+            status = item["status"]
+            for row in matches:
+                row.status = status
+                if status == "unclassified":
+                    row.snippet_id = None
+                    row.reason = None
+                    row.classified_by = None
+                    row.classified_at = None
+                else:
+                    row.snippet_id = (
+                        int(item["snippet_id"]) if status in _NEEDS_TARGET else None
+                    )
+                    row.reason = (item.get("reason") or "").strip() or None
+                    row.classified_by = via
+                    row.classified_at = now
+                classified += 1
+        await session.commit()
+    return {"classified": classified, "unmatched": unmatched}
+
+
+async def list_project_shapes(
+    user_id: int,
+    project_id: int,
+    *,
+    status: str = "",
+    path: str = "",
+    snippet_id: int = 0,
+    include_vanished: bool = False,
+    limit: int = 100,
+    offset: int = 0,
+) -> tuple[list[CodeShape], int]:
+    """A filtered page of a project's ledger, with the unfiltered-match total.
+
+    ([], 0) when the caller can't read the project — the same silence every
+    other project list gives. ``path`` matches the exact file or anything
+    beneath it, mirroring recorded-location semantics.
+    """
+    from sqlalchemy import func, or_
+
+    from scribe.services import access
+
+    if not await access.can_read_project(user_id, project_id):
+        return [], 0
+    conds = [CodeShape.project_id == project_id]
+    if not include_vanished:
+        conds.append(CodeShape.vanished_at.is_(None))
+    if status:
+        conds.append(CodeShape.status == status)
+    if path:
+        clean = path.strip().strip("/")
+        conds.append(or_(
+            CodeShape.path == clean, CodeShape.path.like(clean + "/%")
+        ))
+    if snippet_id:
+        conds.append(CodeShape.snippet_id == snippet_id)
+    async with async_session() as session:
+        total = (
+            await session.execute(
+                select(func.count()).select_from(CodeShape).where(*conds)
+            )
+        ).scalar_one()
+        rows = (
+            await session.execute(
+                select(CodeShape).where(*conds)
+                .order_by(CodeShape.path, CodeShape.symbol, CodeShape.kind)
+                .limit(max(1, min(limit, 500))).offset(max(0, offset))
+            )
+        ).scalars().all()
+    return list(rows), int(total)
+
+
+def _consumer_dict(row: CodeShape) -> dict:
+    """The compact shape a snippet's consumer map carries — enough to open
+    the file, none of the ledger bookkeeping."""
+    out = {
+        "project_id": row.project_id,
+        "repo": row.repo_key,
+        "path": row.path,
+        "symbol": row.symbol,
+        "kind": row.kind,
+        "classified_by": row.classified_by,
+    }
+    if row.reason:
+        out["reason"] = row.reason
+    return out
+
+
+async def snippet_consumers(user_id: int, note_id: int) -> dict:
+    """The structured consumer map for one snippet (#2789): its `instances`
+    (rows judged to conform) and `variants` (named departures, each carrying
+    its why). Rows are filtered to projects the CALLER can read — a shared
+    snippet must not become a side channel into someone else's project
+    layout. Empty lists mean "attach nothing" (#2483)."""
+    from scribe.services import access
+
+    async with async_session() as session:
+        rows = (
+            await session.execute(
+                select(CodeShape).where(
+                    CodeShape.snippet_id == note_id,
+                    CodeShape.status.in_(("instance", "variant")),
+                    CodeShape.vanished_at.is_(None),
+                )
+            )
+        ).scalars().all()
+    readable: dict[int, bool] = {}
+    out: dict[str, list[dict]] = {"instances": [], "variants": []}
+    for row in rows:
+        if row.project_id not in readable:
+            readable[row.project_id] = await access.can_read_project(
+                user_id, row.project_id
+            )
+        if not readable[row.project_id]:
+            continue
+        out["instances" if row.status == "instance" else "variants"].append(
+            _consumer_dict(row)
+        )
+    return out

tests/test_integration_shape_classify.py

@@ -0,0 +1,221 @@
+"""Real-Postgres integration tests for shape classification (#2789).
+
+What mocks can't prove: the all-or-nothing batch against real rows, the
+write-ACL gate, the todo query's filters, and the consumer map riding
+get_snippet. Ledger rows are seeded through the same sync the coverage walk
+uses — no forge needed, the sync takes extracted shapes directly.
+"""
+import pytest
+import pytest_asyncio
+from sqlalchemy import select
+
+from scribe.models import async_session, engine
+from scribe.models.code_shape import CodeShape
+from scribe.models.project import Project
+from scribe.models.user import User
+from scribe.services.shape_ledger import (
+    classify_shapes,
+    list_project_shapes,
+    snippet_consumers,
+    sync_repo_shapes,
+)
+
+pytestmark = pytest.mark.integration
+
+REPO = "git.example.com/alice/widget"
+SHAPES = [
+    ("src/app.py", "sym", "make_app"),
+    ("src/app.py", "sym", "Config"),
+    ("src/util.py", "sym", "helper"),
+    ("web/button.css", "css", "btn"),
+]
+
+
+@pytest_asyncio.fixture(autouse=True)
+async def _dispose_engine():
+    """Dispose the app's module-level engine after each test.
+
+    The engine pools asyncpg connections per event loop, but pytest-asyncio runs
+    each test on a fresh loop — so without this, test 2 gets handed test 1's
+    connection bound to a now-dead loop ("Future attached to a different loop").
+    Disposing in the test's own loop teardown clears the pool cleanly.
+    """
+    yield
+    await engine.dispose()
+
+
+async def _user(session, username: str) -> User:
+    existing = (
+        await session.execute(select(User).where(User.username == username))
+    ).scalar_one_or_none()
+    if existing is not None:
+        return existing
+    user = User(username=username)
+    session.add(user)
+    await session.flush()
+    return user
+
+
+@pytest_asyncio.fixture
+async def seeded():
+    """Owner + outsider, a project with a synced 4-shape ledger, one snippet."""
+    from scribe.services import snippets as snippets_svc
+
+    async with async_session() as s:
+        owner = await _user(s, "classify_owner")
+        other = await _user(s, "classify_other")
+        project = Project(user_id=owner.id, title="Classify target")
+        s.add(project)
+        await s.flush()
+        ids = {"owner": owner.id, "other": other.id, "pid": project.id}
+        await s.commit()
+
+    await sync_repo_shapes(ids["pid"], REPO, SHAPES, seen_marker="main")
+    snippet = await snippets_svc.create_snippet(
+        ids["owner"], name="cls_make_app", code="def make_app():\n    pass\n",
+        language="python", repo="Widget", path="src/factory.py",
+        symbol="factory", project_id=ids["pid"],
+    )
+    ids["snippet"] = int(snippet.id)
+    return ids
+
+
+@pytest.mark.integration
+async def test_classify_applies_judgments_and_reports_unmatched(seeded):
+    owner, pid, sid = seeded["owner"], seeded["pid"], seeded["snippet"]
+    out = await classify_shapes(owner, pid, [
+        {"path": "src/app.py", "symbol": "make_app", "status": "instance",
+         "snippet_id": sid},
+        {"path": "src/util.py", "symbol": "helper", "status": "exempt",
+         "reason": "test scaffolding, deliberately local"},
+        {"path": "web/button.css", "symbol": "btn", "status": "variant",
+         "snippet_id": sid, "reason": "darker focus ring for the toolbar"},
+        {"path": "gone.py", "symbol": "nothing", "status": "exempt",
+         "reason": "x"},
+    ], via="audit")
+    assert out["classified"] == 3
+    assert out["unmatched"] == [{"path": "gone.py", "symbol": "nothing"}]
+
+    rows, total = await list_project_shapes(owner, pid)
+    by_symbol = {r.symbol: r for r in rows}
+    assert total == 4
+    assert by_symbol["make_app"].status == "instance"
+    assert by_symbol["make_app"].snippet_id == sid
+    assert by_symbol["make_app"].classified_by == "audit"
+    assert by_symbol["helper"].status == "exempt"
+    assert by_symbol["helper"].reason == "test scaffolding, deliberately local"
+    assert by_symbol["btn"].status == "variant"
+    assert by_symbol["Config"].status == "unclassified"
+
+    # Withdrawing a judgment returns the shape to the todo, fields cleared.
+    await classify_shapes(owner, pid, [
+        {"path": "src/app.py", "symbol": "make_app", "status": "unclassified"},
+    ])
+    rows, _ = await list_project_shapes(owner, pid, status="unclassified")
+    assert {r.symbol for r in rows} == {"Config", "make_app"}
+    make_app = next(r for r in rows if r.symbol == "make_app")
+    assert make_app.snippet_id is None and make_app.classified_by is None
+
+
+@pytest.mark.integration
+async def test_a_bad_batch_applies_nothing(seeded):
+    """All-or-nothing (#2709's lesson): a caller must never learn later that
+    half a batch silently happened."""
+    owner, pid, sid = seeded["owner"], seeded["pid"], seeded["snippet"]
+    with pytest.raises(ValueError) as err:
+        await classify_shapes(owner, pid, [
+            {"path": "src/app.py", "symbol": "make_app", "status": "instance",
+             "snippet_id": sid},
+            {"path": "src/util.py", "symbol": "helper", "status": "variant",
+             "snippet_id": sid},  # variant with no reason: structural error
+        ])
+    assert "needs a reason" in str(err.value)
+    rows, _ = await list_project_shapes(owner, pid, status="unclassified")
+    assert len(rows) == 4  # including make_app — the valid half did NOT apply
+
+    # A snippet target the caller can't read is the same: nothing applies.
+    with pytest.raises(ValueError):
+        await classify_shapes(owner, pid, [
+            {"path": "src/app.py", "symbol": "make_app", "status": "instance",
+             "snippet_id": 999999999},
+        ])
+
+
+@pytest.mark.integration
+async def test_classification_is_write_gated_and_listing_read_gated(seeded):
+    other, pid, sid = seeded["other"], seeded["pid"], seeded["snippet"]
+    with pytest.raises(ValueError):
+        await classify_shapes(other, pid, [
+            {"path": "src/app.py", "symbol": "make_app", "status": "exempt",
+             "reason": "not their call to make"},
+        ])
+    assert await list_project_shapes(other, pid) == ([], 0)
+
+
+@pytest.mark.integration
+async def test_list_filters_compose(seeded):
+    owner, pid, sid = seeded["owner"], seeded["pid"], seeded["snippet"]
+    await classify_shapes(owner, pid, [
+        {"path": "src/app.py", "symbol": "make_app", "status": "instance",
+         "snippet_id": sid},
+    ])
+    rows, total = await list_project_shapes(owner, pid, path="src")
+    assert total == 3 and all(r.path.startswith("src/") for r in rows)
+    # Directory semantics, not string prefix: "sr" matches nothing.
+    assert (await list_project_shapes(owner, pid, path="sr"))[1] == 0
+    rows, total = await list_project_shapes(owner, pid, snippet_id=sid)
+    assert total == 1 and rows[0].symbol == "make_app"
+    rows, total = await list_project_shapes(
+        owner, pid, status="unclassified", limit=2
+    )
+    assert total == 3 and len(rows) == 2  # paged, with the true total
+
+
+@pytest.mark.integration
+async def test_get_snippet_carries_the_structured_consumer_map(seeded):
+    from scribe.mcp._context import _user_id_ctx
+    from scribe.mcp.tools.snippets import get_snippet
+
+    owner, pid, sid = seeded["owner"], seeded["pid"], seeded["snippet"]
+    await classify_shapes(owner, pid, [
+        {"path": "src/app.py", "symbol": "make_app", "status": "instance",
+         "snippet_id": sid},
+        {"path": "web/button.css", "symbol": "btn", "status": "variant",
+         "snippet_id": sid, "reason": "darker focus ring"},
+    ])
+    token = _user_id_ctx.set(owner)
+    try:
+        data = await get_snippet(snippet_id=sid)
+    finally:
+        _user_id_ctx.reset(token)
+    assert [i["path"] for i in data["instances"]] == ["src/app.py"]
+    assert data["variants"][0]["reason"] == "darker focus ring"
+
+    # The map is caller-scoped: an outsider asking the service directly gets
+    # silence, not another project's file layout.
+    consumers = await snippet_consumers(seeded["other"], sid)
+    assert consumers == {"instances": [], "variants": []}
+
+
+@pytest.mark.integration
+async def test_sync_refiles_rows_whose_snippet_was_purged(seeded):
+    """The SET NULL companion (#2787): a judgment whose target is hard-deleted
+    rejoins the todo on the next sync instead of dangling target-less."""
+    owner, pid, sid = seeded["owner"], seeded["pid"], seeded["snippet"]
+    await classify_shapes(owner, pid, [
+        {"path": "src/app.py", "symbol": "make_app", "status": "instance",
+         "snippet_id": sid},
+    ])
+    from scribe.models.note import Note
+
+    async with async_session() as s:
+        note = await s.get(Note, sid)
+        await s.delete(note)  # hard delete, as purge_trash would
+        await s.commit()
+    await sync_repo_shapes(pid, REPO, SHAPES, seen_marker="main")
+    async with async_session() as s:
+        row = (await s.execute(select(CodeShape).where(
+            CodeShape.project_id == pid, CodeShape.symbol == "make_app",
+        ))).scalar_one()
+    assert row.status == "unclassified"
+    assert row.snippet_id is None

tests/test_mcp_tool_projects.py

@@ -285,7 +285,7 @@ def _enter_project_stubs(p):
 async def test_enter_project_carries_the_bootstrap_ask_when_it_fires():
     """The arrival-moment half of #2683: a mature zero-Systems project greets
     the session with the concrete bootstrap ask, before it is deep in a task —
-    the moment "propose a starter vocabulary" is cheapest."""
+    the moment minting a starter vocabulary is cheapest."""
     import contextlib
 
     ask = "This project has 282 records and NO Systems modelled — ..."

tests/test_mcp_tool_systems.py

@@ -129,7 +129,11 @@ async def test_untagged_hint_escalates_in_a_mature_zero_systems_project():
     assert "282 records" in hint
     assert "Fix scrape retry backoff" in hint      # the project's own evidence
     assert "3-6" in hint and "create_system" in hint
-    assert "propose" in hint
+    # NOT an approval flow (#2798): the agent mints directly, and the
+    # standard cross-project names carry the consistency instead.
+    assert "without asking permission" in hint
+    assert "propose" not in hint and "confirmed" not in hint
+    assert "CI & Release" in hint and "Auth & Access" in hint
     # The generic wording is REPLACED, not appended — two questions is noise.
     assert "no Systems yet" not in hint
 

tests/test_pattern_coverage.py

@@ -19,10 +19,10 @@ from scribe.services.coverage import (
     coverage_line,
     extract_shapes,
     largest_gaps,
-    match_shapes,
     scannable,
     shapes_from_archive,
 )
+from scribe.services.shape_ledger import location_covers
 
 # --- unit: the definition extractor (shared vectors with the hook) -----------
 
@@ -117,51 +117,53 @@ def test_shapes_from_archive_strips_the_wrapper_and_gates_files():
     assert shapes_from_archive(_tarball(TREE)) == TREE_SHAPES
 
 
-# --- unit: matching shapes against recorded locations ------------------------
+# --- unit: the covering predicate (lives with the ledger since #2788) --------
 
 
-def test_match_covers_by_exact_path_dir_prefix_and_css_dot():
+def test_location_covers_by_exact_path_dir_prefix_and_css_dot():
-    recorded = [
+    assert location_covers("src/app.py", "make_app", "src/app.py", "make_app")
-        ("src/app.py", "make_app"),   # exact file
+    # dir prefix + css dot normalization
-        ("web", ".btn"),              # dir prefix + css dot normalization
+    assert location_covers("web", ".btn", "web/button.css", "btn")
-    ]
+    assert not location_covers("src/app.py", "make_app", "src/util.py", "make_app")
-    matched = match_shapes(TREE_SHAPES, recorded)
-    covered = {name for _p, _k, name, ok in matched if ok}
-    assert covered == {"make_app", "btn"}
 
 
 def test_a_symbol_less_record_covers_nothing():
     """A whole-file snippet makes no claim about any particular definition
     inside it — crediting all of them would inflate the number for free."""
-    matched = match_shapes(TREE_SHAPES, [("src/app.py", "")])
+    assert not location_covers("src/app.py", "", "src/app.py", "make_app")
-    assert not any(ok for *_x, ok in matched)
 
 
 def test_no_prefix_bleed_between_sibling_directories():
-    matched = match_shapes(
+    assert not location_covers("src/lib", "helper", "src/library/x.py", "helper")
-        [("src/library/x.py", "sym", "helper")], [("src/lib", "helper")]
-    )
-    assert not matched[0][3]
 
 
-def test_largest_gaps_ranks_by_uncovered_and_drops_clean_dirs():
+def test_largest_gaps_ranks_by_unclassified_and_drops_clean_dirs():
-    matched = match_shapes(TREE_SHAPES, [("src/app.py", "make_app"), ("web", ".btn")])
+    accounted = [
-    gaps = largest_gaps(matched)
+        ("src/app.py", "sym", "make_app", True),
-    assert gaps == [{"dir": "src", "uncovered": 2, "total": 3}]
+        ("src/app.py", "sym", "Config", False),
+        ("src/util.py", "sym", "helper", False),
+        ("web/button.css", "css", "btn", True),
+    ]
+    gaps = largest_gaps(accounted)
+    assert gaps == [{"dir": "src", "unclassified": 2, "total": 3}]
 
 
 def test_coverage_line_is_evidence_carrying_and_labeled_estimate():
     line = coverage_line({
-        "total": 210, "recorded": 34, "estimate": True,
+        "total": 4573, "accounted": 3100, "unclassified": 1473,
+        "counts": {"canonical": 12, "instance": 2900, "variant": 0, "exempt": 188},
+        "estimate": True,
         "computed_at": "2026-08-16T12:00:00+00:00",
         "largest_gaps": [
-            {"dir": "internal/api", "uncovered": 40, "total": 60},
+            {"dir": "internal/api", "unclassified": 40, "total": 60},
-            {"dir": "web/src/components", "uncovered": 25, "total": 30},
+            {"dir": "web/src/components", "unclassified": 25, "total": 30},
         ],
     })
-    assert "34/210 shapes recorded" in line
+    assert "3100/4573 shapes accounted for" in line
+    assert "12 canonical · 2900 instance · 188 exempt" in line  # zero variant elided
     assert "estimate" in line
     assert "2026-08-16" in line
+    assert "1473 unclassified" in line
     assert "internal/api, web/src/components" in line
 
 
@@ -257,19 +259,52 @@ async def test_coverage_measures_the_tree_exactly_and_caches(seeded):
         refresh_coverage,
     )
 
+    from sqlalchemy import select
+
+    from scribe.models import async_session
+    from scribe.models.code_shape import CodeShape
+
     uid, pid = seeded["uid"], seeded["pid"]
     selector = _selector(_tarball(TREE))
 
     coverage = await compute_coverage(uid, pid, selector=selector)
     assert coverage is not None
     assert coverage["total"] == 4
-    assert coverage["recorded"] == 2
+    assert coverage["accounted"] == 2
+    assert coverage["unclassified"] == 2
+    assert coverage["counts"] == {
+        "canonical": 2, "instance": 0, "variant": 0, "exempt": 0,
+    }
     assert coverage["estimate"] is True
     assert coverage["repos"] == [{
         "repo": "git.example.com/alice/widget", "ref": "main",
-        "total": 4, "recorded": 2,
+        "total": 4, "accounted": 2,
     }]
-    assert coverage["largest_gaps"] == [{"dir": "src", "uncovered": 2, "total": 3}]
+    assert coverage["largest_gaps"] == [
+        {"dir": "src", "unclassified": 2, "total": 3}
+    ]
+
+    # The walk fed the LEDGER (#2788): every extracted shape has a row, the
+    # snippet reference locations are mechanically stamped canonical WITH
+    # their snippet id, and the rest sit in the todo state.
+    async with async_session() as s:
+        rows = (await s.execute(
+            select(CodeShape).where(CodeShape.project_id == pid)
+        )).scalars().all()
+    by_symbol = {r.symbol: r for r in rows}
+    assert set(by_symbol) == {"make_app", "Config", "helper", "btn"}
+    assert by_symbol["make_app"].status == "canonical"
+    assert by_symbol["make_app"].snippet_id is not None
+    assert by_symbol["make_app"].classified_by == "mechanical"
+    assert by_symbol["btn"].status == "canonical"
+    assert by_symbol["Config"].status == "unclassified"
+    assert by_symbol["helper"].status == "unclassified"
+    assert all(r.vanished_at is None for r in rows)
+    assert all(r.first_seen_commit for r in rows)  # ref at minimum
+
+    # Idempotence: a second walk changes nothing about the readout.
+    again = await compute_coverage(uid, pid, selector=_selector(_tarball(TREE)))
+    assert (again["total"], again["accounted"]) == (4, 2)
 
     # Nothing computed → nothing cached; refresh writes; the cache reads back
     # byte-equal, because enter_project will serve exactly this.
@@ -278,6 +313,57 @@ async def test_coverage_measures_the_tree_exactly_and_caches(seeded):
     assert (await cached_coverage(uid, pid)) == json.loads(json.dumps(stored))
 
 
+@pytest.mark.integration
+async def test_ledger_keeps_judgments_and_stamps_vanished_shapes(seeded):
+    """The two survival rules (#2788): an agent's classification outlives
+    recompute, and a shape that leaves the tree is stamped vanished — kept
+    for history, dropped from the readout."""
+    from datetime import datetime, timezone
+
+    from sqlalchemy import select
+
+    from scribe.models import async_session
+    from scribe.models.code_shape import CodeShape
+    from scribe.services.coverage import compute_coverage
+
+    uid, pid = seeded["uid"], seeded["pid"]
+    await compute_coverage(uid, pid, selector=_selector(_tarball(TREE)))
+
+    # An agent judges `helper` a deliberate one-off.
+    async with async_session() as s:
+        helper = (await s.execute(select(CodeShape).where(
+            CodeShape.project_id == pid, CodeShape.symbol == "helper",
+        ))).scalar_one()
+        helper.status = "exempt"
+        helper.reason = "test scaffolding, deliberately local"
+        helper.classified_by = "agent"
+        helper.classified_at = datetime.now(timezone.utc)
+        await s.commit()
+
+    # The tree moves on: util.py (helper) is gone entirely, app.py loses
+    # nothing. The judgment on `helper` must survive AS HISTORY (vanished,
+    # still exempt), never be reset by the sync.
+    smaller = {k: v for k, v in TREE.items() if k != "src/util.py"}
+    coverage = await compute_coverage(uid, pid, selector=_selector(_tarball(smaller)))
+    assert coverage["total"] == 3          # helper's row left the readout
+    assert coverage["accounted"] == 2
+    assert coverage["counts"]["exempt"] == 0   # vanished rows don't count
+
+    async with async_session() as s:
+        helper = (await s.execute(select(CodeShape).where(
+            CodeShape.project_id == pid, CodeShape.symbol == "helper",
+        ))).scalar_one()
+    assert helper.vanished_at is not None
+    assert helper.status == "exempt"       # the judgment is history, kept
+    assert helper.reason == "test scaffolding, deliberately local"
+
+    # And it returns: the shape reappearing clears the stamp, judgment intact.
+    coverage = await compute_coverage(uid, pid, selector=_selector(_tarball(TREE)))
+    assert coverage["total"] == 4
+    assert coverage["accounted"] == 3      # exempt counts as accounted again
+    assert coverage["counts"]["exempt"] == 1
+
+
 @pytest.mark.integration
 async def test_enter_project_surfaces_the_line_only_once_computed(seeded):
     from scribe.mcp._context import _user_id_ctx
@@ -296,9 +382,10 @@ async def test_enter_project_surfaces_the_line_only_once_computed(seeded):
         after = await enter_project(project_id=pid)
         line = after["pattern_coverage"]
         assert line.startswith(
-            "pattern-library coverage: 2/4 shapes recorded (estimate, computed "
+            "shape accounting: 2/4 shapes accounted for — 2 canonical "
+            "(estimate, computed "
         )
-        assert line.endswith("; largest gaps: src")
+        assert line.endswith("; 2 unclassified, largest: src")
     finally:
         _user_id_ctx.reset(token)
 

tests/test_services_backup.py

@@ -13,17 +13,19 @@ import pytest
 from scribe.services import backup
 
 
-def test_backup_version_is_v6():
+def test_backup_version_is_v7():
-    """v6 added note_supersessions (#278). The bump is the point of the test —
+    """v7 added code_shapes (#2787). The bump is the point of the test —
     a payload section added without moving the version produces backups that
     are structurally different and indistinguishable by inspection."""
-    assert backup.BACKUP_VERSION == 6
+    assert backup.BACKUP_VERSION == 7
 
 
 def test_not_included_lists_the_known_gaps():
     # The deferred tables must be surfaced explicitly, not silently dropped.
+    # forge_connections is excluded as CREDENTIALS (api_keys reasoning): a
+    # backup that carries forge tokens is a token-exfiltration file (#2778).
     for table in ("groups", "project_shares", "note_shares", "api_keys",
-                  "note_embeddings", "retrieval_logs"):
+                  "note_embeddings", "retrieval_logs", "forge_connections"):
         assert table in backup._NOT_INCLUDED
 
 
@@ -105,14 +107,14 @@ async def test_export_full_backup_contains_every_declared_section():
     assert out["version"] == backup.BACKUP_VERSION
     assert out["scope"] == "full"
     assert "api_keys" in out["_not_included"]
-    # The sections v2 silently dropped, the six v5 added, and v6's
+    # The sections v2 silently dropped, the six v5 added, v6's
-    # note_supersessions (all empty here).
+    # note_supersessions, and v7's code_shapes (all empty here).
     for key in ("rulebooks", "rulebook_topics", "rules",
                 "rulebook_subscriptions", "rule_suppressions",
                 "topic_suppressions",
                 "systems", "record_systems", "design_systems",
                 "design_tokens", "note_usage_events", "repo_bindings",
-                "note_supersessions"):
+                "note_supersessions", "code_shapes"):
         assert key in out, f"missing export section: {key}"
         assert out[key] == []
 

tests/test_shape_ledger.py

@@ -0,0 +1,84 @@
+"""The shape ledger's schema contract (#2787, milestone 294, note 2786).
+
+Step 1 pins the model: identity, the classification vocabulary, and the
+token-free serialisation. The sync pass (step 2) and the classification
+surface (step 3) grow their tests here; DB-backed behavior lands in the
+integration lane once there is behavior to exercise.
+"""
+from scribe.models import Base
+from scribe.models.code_shape import SHAPE_CLASSIFIERS, SHAPE_STATUSES, CodeShape
+
+
+def test_identity_is_project_repo_path_symbol_kind():
+    """Kind is part of identity on purpose: one file can define `.foo` (css)
+    and `foo` (sym) as distinct shapes — the extractor emits both."""
+    table = Base.metadata.tables["code_shapes"]
+    unique = next(
+        c for c in table.constraints
+        if getattr(c, "name", "") == "uq_code_shapes_identity"
+    )
+    assert [c.name for c in unique.columns] == [
+        "project_id", "repo_key", "path", "symbol", "kind",
+    ]
+
+
+def test_the_todo_state_is_the_default():
+    """A shape nobody has judged yet must read `unclassified` — the ledger's
+    todo list — never silently look classified."""
+    assert CodeShape.__table__.c.status.default.arg == "unclassified"
+    assert "unclassified" in SHAPE_STATUSES
+    assert set(SHAPE_STATUSES) == {
+        "canonical", "instance", "variant", "exempt", "unclassified",
+    }
+    assert set(SHAPE_CLASSIFIERS) == {
+        "agent", "audit", "hook", "mechanical", "import",
+    }
+
+
+def test_snippet_reference_survives_snippet_deletion_as_null():
+    """SET NULL, not CASCADE: a deleted snippet must not silently erase the
+    accounting rows that pointed at it — the sync pass re-files them as
+    unclassified so they rejoin the todo."""
+    fk = next(iter(CodeShape.__table__.c.snippet_id.foreign_keys))
+    assert fk.ondelete == "SET NULL"
+    assert fk.column.table.name == "notes"
+
+
+def test_status_queries_have_an_index():
+    """list_shapes(status=unclassified) is THE todo query (step 3) — it must
+    not degrade into a project-wide scan as ledgers reach thousands of rows."""
+    names = {ix.name for ix in CodeShape.__table__.indexes}
+    assert "ix_code_shapes_project_status" in names
+    assert "ix_code_shapes_snippet" in names
+
+
+# --- step 3: the classification batch validator (pure, checked before ACL) ---
+
+
+def test_batch_validation_names_the_failing_item():
+    from scribe.services.shape_ledger import validate_classifications as v
+
+    ok = {"path": "src/a.py", "symbol": "f", "status": "exempt", "reason": "one-off"}
+    assert v([ok]) is None
+    assert "empty" in v([])
+    assert "classifications[1]" in v([ok, {"symbol": "f", "status": "exempt"}])
+    assert "unknown status" in v([{**ok, "status": "covered"}])
+    # A judgment that references canon must name the canon...
+    assert "needs snippet_id" in v(
+        [{"path": "a", "symbol": "f", "status": "instance"}]
+    )
+    # ...and a departure/exemption must carry its why — the why IS the record.
+    assert "needs a reason" in v(
+        [{"path": "a", "symbol": "f", "status": "variant", "snippet_id": 3}]
+    )
+    assert "needs a reason" in v([{"path": "a", "symbol": "f", "status": "exempt"}])
+    # Withdrawing a judgment needs neither target nor reason.
+    assert v([{"path": "a", "symbol": "f", "status": "unclassified"}]) is None
+
+
+def test_classify_and_list_are_mounted_as_mcp_tools():
+    from scribe.mcp.server import build_mcp_server
+
+    mcp = build_mcp_server()
+    for name in ("classify_shapes", "list_shapes"):
+        assert mcp._tool_manager.get_tool(name) is not None