feat(500): the reply shapes become server product content - a compact core and per-kind slices, each tied to its moment (#5493)

services/reply_shapes.py is the single source of the default reply shapes:
the core (every reply, ~1,800 chars against a 2,200 budget) and three
slices - completion on work.finish, asks on reply.ask, plan on work.plan.
Read through list_reply_shapes (MCP, read-only) and GET
/api/retrieval/reply-shapes, one service behind both doors.

Nothing delivers them yet; that is step 3. The skill still carries its
copy until step 4 shrinks it to the long-form reference.

The software-only vocabulary guard moves into tests/helpers.py
(DEV_ONLY, dev_only_hits) rather than becoming a fourth copy.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

src/scribe/services/reply_shapes.py

@@ -0,0 +1,194 @@
+"""The default shapes of a reply, and the moment each one rides (milestone 500).
+
+WHY THIS EXISTS
+
+A reply is where the person the work is for finds out what happened, and they
+were not there while it was done. Milestone 409 wrote the shapes that make a
+reply readable to them — conclusion first, the shortest reply that carries the
+answer, where the work stands — and put them in one bundled skill. A skill
+arrives only when the agent decides to load it, so in a long session the shape
+faded out of context exactly when replies were getting longer.
+
+So the shapes are product content held here, on the server, and delivered by
+the moments they belong to rather than by the agent remembering to ask:
+
+  - the CORE applies to every reply, so it rides every turn — in full once per
+    session and after a compaction, and as a one-line pointer otherwise;
+  - each SLICE belongs to one kind of reply, and arrives at the moment that
+    comes just before that kind is written: closing a task comes before a
+    completion report, a structured question before an ask, opening a plan
+    before a plan is put up for review.
+
+The moment already knows which kind of reply is coming, so there is no "pick
+the type, then load its shape" step for an agent to skip.
+
+ONE COPY
+
+This module is the single source. The `reporting-back` skill keeps the
+long-form reference (a worked example, the reasoning) and points here; the
+hooks carry timing and transport and say nothing of their own about shape. An
+operator's adjustments are `preference` records mounted on the same moments,
+and arrive beside the default — where the two differ, the preference is what
+the operator asked for.
+
+HOW TO WRITE ONE
+
+Domain-neutral (rule 115): the work may be software, configuration,
+infrastructure or anything else a person drives through an agent. "Evidence"
+is what the reader could open to check; "verified" is whatever checking means
+for that work. The test beside this module refuses software-only vocabulary.
+
+Short, because it is paid for in every session's context. The core's budget is
+asserted; a sentence added to it should replace one.
+"""
+from __future__ import annotations
+
+from dataclasses import dataclass
+
+from scribe.services import moments
+
+
+@dataclass(frozen=True)
+class ReplyShape:
+    """One piece of the default reply shape, and when it arrives."""
+
+    key: str
+    """Stable identifier: what a delivery records and the Settings view keys on."""
+
+    title: str
+    """What the piece is, as the Settings view names it."""
+
+    moment: str
+    """The catalog moment this piece rides — where an operator's own
+    preferences for it are mounted too."""
+
+    delivered: str
+    """When it arrives, said plainly for the Settings view."""
+
+    text: str
+    """The shape itself, as the agent reads it."""
+
+
+CORE_KEY = "core"
+
+# ~550 tokens at four characters a token. The core is paid for in every
+# session, so this is the ceiling the test holds it to — not a target. It
+# shipped at ~1,800, leaving room for the one-line judgement (milestone 500
+# step 2) and little else.
+CORE_BUDGET_CHARS = 2200
+# A slice is paid for only at its moment, and once per session; it can afford
+# more than the core and still has to stay a slice, not the old skill again.
+SLICE_BUDGET_CHARS = 1400
+
+
+_CORE = """\
+Shape the reply around where the work stands, not the order you did things in. \
+The reader was not there while you worked.
+
+- **Conclusion first** — the result, the verdict or the question, then what supports it.
+- **The shortest reply that carries the answer.** Length is work handed back to the \
+reader. It is earned by a comparison they asked for, options that need laying side by \
+side, or numbers that are the point.
+- **One topic per section; bold the few things that matter.**
+- **Plain words** — the reader's vocabulary, not names you coined while working.
+- **Place the work** — the task or plan it belongs to, by id and title, read from the \
+record rather than recalled. Work with no record behind it says so.
+- **Answer the standing questions, even with "nothing"**: does anything need them, and \
+what happens next. A section that explains earns its place only when it changes what \
+they do or decide; the rest belongs in the record's log.
+- **A decision they have made is the input to the work.** Act on it; reopen it only \
+for new evidence, said once.
+- **Holding an action until they say yes?** Put it near the top, under "Approval requested".
+- **End with the one thing they need to decide or do, in bold** — or say there is none.
+
+By kind:
+- **Answer** — the answer, then what they could open to check it. An evaluation is \
+verdict · what exists · the gaps · a recommendation.
+- **Proposal** — two or three approaches, the trade-off of each, one recommendation. \
+A review is findings ranked by how much they matter.
+- **Progress** — a line or two. **Blocked** — what stopped, what you tried, what you need.
+- **Where are we** — the plan and its progress · done · open · needs you · next.
+
+Before sending, read it as someone who was not there, reading quickly; then once more \
+for what can go."""
+
+
+_COMPLETION = """\
+A completion report, from the `placement` the closing call returned:
+
+- **Where this sits** — the plan and the step, or the task alone when it has no plan.
+- **What now works** — outcomes the reader would notice ("you can now…", "X no longer…"), \
+not the steps behind them; those belong in the task's log.
+- **How / why** — only the decisions worth knowing, and how it was verified. Anything \
+that could not be verified is said here rather than left to read as passed.
+- **Needs you** — an action, an approval, a decision, or "nothing". It takes two tests: \
+is it theirs to decide, and is work waiting on it? A question you could settle by \
+reading or measuring something is work not yet done — settle it, say which way you went, \
+and leave them free to overrule.
+- **Next** — from `placement.next`, or say the plan is finished. An offer to fix \
+something you found goes here."""
+
+
+_ASKS = """\
+Asking them to decide or to act — pick the shape:
+
+- **Decision** — the question first · two to four options, each with what it changes · \
+your recommendation first among them.
+- **Clarification** — "My reading is X · the gap is Y · unless you say otherwise I'll do Z."
+- **Handoff** (only they can do it) — the action · why it needs them · what it unblocks · \
+what you will do after.
+- **Approval** (ready, and holding for a yes) — under "Approval requested": exactly what \
+happens on yes, one numbered item per change so they can approve part · why it needs \
+them · how it is undone.
+- **Conflict** (what you are about to do clashes with a rule, a plan or an earlier \
+decision) — what it says · what you were about to do · where they clash · A or B?
+
+Before asking, check whether you can find the answer yourself: something you can read \
+or measure is a fact to check, not a question to send. When an option rests on existing \
+behaviour, say whose call that was — theirs (cite it) or a past session's nobody confirmed."""
+
+
+_PLAN = """\
+A plan put up for review, before starting:
+
+- **Goal** — what done looks like, and why, in a sentence or two.
+- **Steps** — in order, each small enough to check on its own.
+- **How you will know it works** — something observable, not "it is written".
+- **Open questions** — only the ones that are theirs to answer."""
+
+
+SHAPES: dict[str, ReplyShape] = {s.key: s for s in (
+    ReplyShape(CORE_KEY, "Every reply", "reply.report",
+               "every turn — in full once per session and after a compaction, "
+               "otherwise as a one-line pointer", _CORE),
+    ReplyShape("completion", "Completion report", "work.finish",
+               "when a piece of work is closed", _COMPLETION),
+    ReplyShape("asks", "Asking them to decide or act", "reply.ask",
+               "when a structured question is put to them", _ASKS),
+    ReplyShape("plan", "A plan for review", "work.plan",
+               "when a plan is opened", _PLAN),
+)}
+
+
+def core() -> ReplyShape:
+    return SHAPES[CORE_KEY]
+
+
+def for_moments(names: list[str]) -> list[ReplyShape]:
+    """The slices that ride these moments, in catalog order. The core is not a
+    slice: it rides every turn, whatever moment the turn reaches."""
+    wanted = set(names)
+    return [s for s in SHAPES.values() if s.key != CORE_KEY and s.moment in wanted]
+
+
+def catalog() -> dict:
+    """The shapes as plain data, for the MCP tool and the REST door alike."""
+    return {
+        "shapes": [
+            {"key": s.key, "title": s.title, "moment": s.moment,
+             "means": moments.MOMENTS[s.moment].means,
+             "delivered": s.delivered, "text": s.text}
+            for s in SHAPES.values()
+        ],
+        "total": len(SHAPES),
+    }