docs(6205): retrieval_pipeline comments state the reason, not the instance that taught it

Record ids, milestone and step numbers, and the stories of copies drifting
apart are removed; each invariant and slot is justified in present tense.
The rule-band probe table stays as calibration. Section headers drop their
milestone tags.

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

src/scribe/services/retrieval_pipeline.py

@@ -1,30 +1,26 @@
 """One retrieval pipeline: a ranked surface is a spec, not a hand-copied arm.
 
-Milestone 456. Every ranked rule surface used to write out the same sequence
-by hand — search, band, split fresh from repeats, log the call before any early
-return, reserve a slot, render, record what was surfaced — once per arm, each
-modelled on the nearest sibling. The copies drifted, and the defects that
-mattered were fixed one copy at a time (#3497, #3750, #3752): "this arm shipped
-with the same defect inherited from its sibling" was a comment in the code.
-The invariants were held together by a parametrised test over three copies
-rather than by there being one implementation.
+Every ranked rule surface runs the same sequence — search, band, split fresh
+from repeats, log the call before any early return, reserve a slot, render,
+record what was surfaced. Written out by hand once per arm, each modelled on
+the nearest sibling, the copies drift, and a defect gets fixed in one copy
+and inherited by the next.
 
 Here the sequence is written ONCE (`run_rule_arm`), and an arm is a `RuleArm`:
 which source it records under, and which of the stages it takes. The
-differences between arms that used to be buried in their copies are now four
-booleans on one screen — and the ones that are accidents rather than
-decisions are milestone 456 step 7's to settle, one at a time, on evidence.
+differences between arms are four booleans on one screen, where an accident
+is visible as one.
 
 THE INVARIANTS, each written once:
 
-  - the call row is written before any early return (#3497), because a call
-    that found nothing is the only evidence a bar is too high;
+  - the call row is written before any early return, because a call that
+    found nothing is the only evidence a bar is too high;
   - the call row's results and the surfacing rows are both the FRESH cut, so
-    the two tables describe the same delivery (#3668, #3752);
+    the two tables describe the same delivery;
   - the session ledger decides how a repeat RENDERS, never whether it is
-    retrieved (#3750, #4101);
-  - band before dedup (#3851), so "you were shown this" never changes what
-    counts as relevant;
+    retrieved;
+  - band before dedup, so "you were shown this" never changes what counts
+    as relevant;
   - telemetry never costs the reader a line, and the arm as a whole fails
     open: a recall aid may never break the prompt, the command or the write.
 
@@ -43,14 +39,14 @@ from scribe.services.lessons import LESSON_NOTE_TYPE, claim_line
 
 logger = logging.getLogger(__name__)
 
-# ── Shared rendering and ranking helpers (moved from plugin_context) ──────
+# ── Shared rendering and ranking helpers ─────────────────────────────────
 
-# MEASURED, NOT REASONED — and the reasoning it replaced was wrong (#3851).
+# MEASURED, NOT REASONED.
 #
-# The prediction was that rules would rank SHARPLY, because `rule_document()`
+# The natural prediction is that rules rank SHARPLY, because `rule_document()`
 # shapes them the way snippets are shaped — trigger in the embedded title and
-# again above the body — and note 2485 measured snippets separating their top
-# hit by 0.153 while every other kind managed 0.010–0.023.
+# again above the body — and snippets separate their top hit by ~0.15 while
+# every other kind manages 0.010–0.023.
 #
 # They do not. Three probes against real act queries, scored by the same
 # embedding the arms use:
@@ -60,21 +56,20 @@ logger = logging.getLogger(__name__)
 #   a bare-owner-filter query     top 0.656, gap to second 0.020
 #
 # That is dev-log territory, not snippet territory: rules arrive as a
-# tightly-packed block. Shaping alone did not buy separation, which is worth
-# recording because the opposite was the natural inference from 2485.
+# tightly-packed block. Shaping alone does not buy separation.
 #
 # So the band is narrow BECAUSE the corpus is flat. At 0.10 — the notes
 # menu's value — every one of the top eight on the push probe falls inside,
 # including a CI-registry rule and another project's branch policy. At 0.05
 # it admits roughly three ranks, which is the span where the scores are still
-# saying something. 2485's Finding 3 is the standing caveat: no band value
-# fixes a tie, and if rules ever rank as flat as dev-logs did this control
-# stops working and the answer is a reranker (#1038), not a smaller number.
+# saying something. The standing caveat: no band value fixes a tie, and if
+# rules ever rank as flat as dev-logs this control stops working and the
+# answer is a reranker, not a smaller number.
 _RULEHINT_BAND = 0.05
 
 
 def _rule_band(hits: list) -> list:
-    """The top hit, plus every hit within `_RULEHINT_BAND` of it (#3851).
+    """The top hit, plus every hit within `_RULEHINT_BAND` of it.
 
     The instrument that lets an act surface a SET without inventing one: a
     fixed k fills its slots whether or not anything deserves them, while this
@@ -82,8 +77,8 @@ def _rule_band(hits: list) -> list:
     still shows one and four competing rules show four.
 
     Sync and pure, and deliberately its own function rather than a comparison
-    written twice — the two act arms are the pair #3497 records drifting apart
-    by being modelled on each other instead of sharing.
+    written twice: the two act arms drift apart when each is modelled on the
+    other instead of sharing.
 
     Takes `(score, rule)` pairs already ordered best-first, as both
     `semantic_search_*` helpers return them.
@@ -100,9 +95,7 @@ def checkpoint_for(
     """The one rule, if any, that should STOP this act rather than annotate it.
 
     Sync and pure so it can be read against a fixed list of hits without a
-    database — the two arms share it for the reason `_rule_band` is shared:
-    #3497 is the record of these two drifting apart by being modelled on each
-    other instead of sharing one function.
+    database — the two arms share it for the reason `_rule_band` is shared.
 
     FOUR CONDITIONS, AND EACH IS A DIFFERENT KIND OF WRONG IT PREVENTS.
 
@@ -119,7 +112,7 @@ def checkpoint_for(
        names it.
 
     3. THE SESSION HAS NOT OPENED IT. `held` is observable — a PostToolUse
-       hook watches for the `get_rule` call (#4100) — so this is a recorded
+       hook watches for the `get_rule` call — so this is a recorded
        event and not a model's self-report about its own context. A session
        that read the rule has already had the thing the checkpoint exists to
        produce, and stopping it again would be punishing the behaviour being
@@ -131,9 +124,8 @@ def checkpoint_for(
        a stop justified by a score nobody claimed.
 
     Returns a dict rather than a rule or a tuple. Widening a tuple is an
-    interface change to every unpack site that the compiler does not report
-    (#4207, learned the expensive way in this same milestone's first week), and
-    this value crosses a JSON boundary into a shell script where a missing
+    interface change to every unpack site that the compiler does not report,
+    and this value crosses a JSON boundary into a shell script where a missing
     field is a silently empty variable.
     """
     if not kept or floor <= 0:
@@ -154,7 +146,7 @@ def checkpoint_for(
     }
     # RENDERED HERE, at the one call site, rather than by each arm. Two arms
     # that each remember to render it are two arms that can stop agreeing on
-    # what a stop says — which is #3497's history for this exact pair. The
+    # what a stop says. The
     # text stays a separate function so it can be read and tested without a
     # rule object, but nothing outside this line decides whether to call it.
     found["reason"] = checkpoint_reason(found)
@@ -164,7 +156,7 @@ def checkpoint_for(
 def checkpoint_reason(checkpoint: dict) -> str:
     """The text the agent reads INSTEAD of running the act.
 
-    Written as a practice rather than a prohibition (rule 165): it says what
+    Written as a practice rather than a prohibition: it says what
     to do and why it is worth doing, not what is forbidden. The act is not
     wrong — nothing here knows whether it is — and saying so plainly is what
     keeps the stop from reading as an accusation the system is in no position
@@ -195,10 +187,10 @@ def checkpoint_reason(checkpoint: dict) -> str:
 def _rule_hint_line(
     rule, *, where: str, seen: bool, held: bool = False, compact: bool = False,
 ) -> str:
-    """One rule hint line — both arms, both tails, both kinds (#3750, #3849).
+    """One rule hint line — both arms, both tails, both kinds.
 
-    THREE INDEPENDENT AXES SINCE #3851. `compact` joins `kind` and `seen`, and
-    like them it reads none of the others: it says how much ROOM this line
+    THREE INDEPENDENT AXES. `compact` joins `kind` and `seen`, and like them
+    it reads none of the others: it says how much ROOM this line
     gets, which is a fact about its rank among today's hits rather than about
     the rule. A compact line is still a full claim that the rule may apply —
     it simply cites the rule instead of quoting its trigger. Crucially it
@@ -206,13 +198,11 @@ def _rule_hint_line(
     independent: shortening a line must not decide what it says about whether
     the session is holding the rule.
 
-    WHY THE LATER LINES ARE QUIETER. Measured at #3851: a full line runs ~143
-    tokens once the trigger is rendered, and #3855 tripled trigger lengths
-    across the corpus, so five full lines cost ~568 tokens before every Bash
-    call. Top-full-plus-references costs ~299 — about 2x the old single line,
-    for four more rules. The budget argument that once justified a single slot
-    was real; what it actually forbids is four voices at full volume, not four
-    voices.
+    WHY THE LATER LINES ARE QUIETER. A full line runs ~143 tokens once the
+    trigger is rendered, so five full lines cost ~568 tokens before every Bash
+    call. Top-full-plus-references costs ~299 — about 2x a single full line,
+    for four more rules. The token budget forbids four voices at full volume,
+    not four voices.
 
     The top hit keeps the full rendering because it is the one the ranker is
     most confident about, and a reader who acts on exactly one line should
@@ -222,32 +212,31 @@ def _rule_hint_line(
     and neither reads the other. A preference and a rule differ in force; a
     repeat and a first surfacing differ in whether the session already holds
     the line. Those are unrelated facts, and keeping them unrelated in the
-    code is what stopped the second kind from reopening the repeat question.
+    code is what lets a new kind join without reopening the repeat question.
 
     ONE FUNCTION BECAUSE THE TAILS MUST NOT DRIFT. The two arms phrase their
     heads differently ("may apply here" vs "may apply to this Bash call") and
-    that difference is deliberate. Everything after it must not differ, and
-    #3497's history is that the pre-tool arm inherited a defect from its
-    sibling by being modelled on it rather than sharing with it. Two copies of
-    a two-branch string is how one branch gets fixed and the other does not.
+    that difference is deliberate. Everything after it must not differ: an
+    arm modelled on its sibling rather than sharing with it inherits its
+    defects, and two copies of a two-branch string is how one branch gets
+    fixed and the other does not.
 
-    WHY A REPEAT GETS A LINE AT ALL. Both arms used to drop a hit whose id was
-    already on the session's exclusion ledger and emit nothing. That is correct
-    only while the session still HOLDS what it was told, and a compaction
-    breaks exactly that: the earlier injection is summarized away while the id
-    stays on the ledger, so the rule is absent from context AND unreachable for
-    the rest of the session (#3749 closes the compaction half; this closes the
-    ordinary half, where a session simply stops holding a line it read an hour
-    ago).
+    WHY A REPEAT GETS A LINE AT ALL. Dropping a hit whose id is already on the
+    session's exclusion ledger is correct only while the session still HOLDS
+    what it was told, and a compaction breaks exactly that: the earlier
+    injection is summarized away while the id stays on the ledger, so the rule
+    would be absent from context AND unreachable for the rest of the session.
+    The same holds, more slowly, for a session that simply stops holding a
+    line it read an hour ago.
 
-    Only ONE CLAUSE of the original line is false on a repeat — the claim that
+    Only ONE CLAUSE of a first-surfacing line is false on a repeat — the claim that
     the rule is not in the session's loaded set. So only that clause changes.
     Title, trigger and pull pointer are identical either way, the statement is
     never injected either way, and a repeat therefore costs the same ~40 tokens
     as a first surfacing and no more.
 
     DELIBERATELY NOT ASKING THE SESSION WHETHER IT HOLDS THE RULE. A model
-    asked "do you still hold rule 156?" will say yes, and the claim is
+    asked "do you still hold rule N?" will say yes, and the claim is
     unverifiable self-report about its own context. The answer is also not
     needed: the line is cheap enough to always emit and carries its own remedy
     in both branches. Removing the question removes the fragility rather than
@@ -265,8 +254,8 @@ def _rule_hint_line(
     # A reader skimming an injected block gets one word to place the register \u2014
     # so it is the SECOND word that moves, and it is the word naming force.
     # Everything structural after it is identical, so the three kinds read as
-    # one set rather than three formats (milestone 385's step 5 writes the
-    # lesson voice against these two; they are one paragraph, not three).
+    # one set rather than three formats (the lesson voice is written against
+    # these two; they are one paragraph, not three).
     noun = "Preference" if preference else "Standing rule"
     # The only other place force is asserted. A rule's line tells the reader
     # not to dismiss it unread, because dismissing a rule unread is how the
@@ -277,21 +266,21 @@ def _rule_hint_line(
         "for how this has been done before" if preference
         else "before deciding it does not apply"
     )
-    # THREE STATES, BECAUSE TWO OF THEM WERE BEING TOLD THE SAME LIE (#4100).
+    # THREE STATES, BECAUSE TWO OF THEM MUST NOT BE TOLD THE SAME THING.
     #
     # `seen` means an arm NAMED this rule earlier. It does not mean the session
     # read it — the line is a teaser, and a teaser skimmed past leaves nothing
     # behind, least of all after a compaction summarises the turn it arrived
-    # in. "You saw it earlier this session" asserted something about the
-    # reader's context that the server had no way to know.
+    # in. "You saw it earlier this session" would assert something about the
+    # reader's context that the server has no way to know.
     #
     # `held` is the observable half: a PostToolUse hook watches for the
     # `get_rule` call itself, so this is a recorded EVENT rather than a claim.
     # That distinction is what keeps the non-goal above intact — the objection
     # was to asking a model about its own context, not to noticing what it did.
     #
-    # The middle state is the honest one and the one that was missing: named,
-    # not opened. It gets the full invitation, because a session that skipped
+    # The middle state is the honest one: named, not opened. It gets the full
+    # invitation, because a session that skipped
     # the teaser is in almost the same position as one that never saw it.
     if held:
         tail = (
@@ -310,17 +299,17 @@ def _rule_hint_line(
         )
     if compact:
         # THE TRIGGER GOES; THE TAIL STAYS. Only one of the two is expensive \u2014
-        # a trigger runs 300-400 characters after #3855, the tail about 100 \u2014
+        # a trigger runs 300-400 characters, the tail about 100 \u2014
         # so dropping the trigger is nearly the whole saving and dropping the
         # tail would be mostly sacrifice.
         #
-        # It would also destroy the one thing #3750 exists to say. The tail is
-        # what tells a reader whether they were already told this and may no
-        # longer be holding it, and that is the entire difference they can act
-        # on; a reference with no tail reads as a first surfacing whether it is
-        # one or not. This branch shipped without it for one commit and
+        # It would also destroy the one thing a repeat line exists to say. The
+        # tail is what tells a reader whether they were already told this and
+        # may no longer be holding it, and that is the entire difference they
+        # can act on; a reference with no tail reads as a first surfacing
+        # whether it is one or not.
         # test_a_rule_the_session_already_holds_is_referenced_not_re_offered
-        # caught it, which is the guard working exactly as #3750 intended.
+        # pins this.
         return (
             f"Also \u2014 {noun.lower()} \u201c{rule.title}\u201d. {tail}"
         )
@@ -331,14 +320,14 @@ def _rule_hint_line(
     )
 
 
-# ── What a spec declares about itself (milestone 456 step 5) ─────────────
+# ── What a spec declares about itself ────────────────────────────────────
 #
-# A ranked surface used to be described in three hand-kept lists besides its
-# code: its tuning pair in `retrieval_surfaces.SURFACES`, its row in
+# A ranked surface is described in three places besides its code: its tuning
+# pair in `retrieval_surfaces.SURFACES`, its row in
 # `retrieval_registry.POINTS`, and — for a rule surface — its membership in
-# `rule_usage.RANKED_SOURCES`. Tests existed to make the four agree. Now the
-# spec carries all of it, and the three lists are read off the specs, so an
-# arm added here is tunable, measured and counted by construction.
+# `rule_usage.RANKED_SOURCES`. The spec carries all of it, and the three lists
+# are read off the specs, so an arm added here is tunable, measured and
+# counted by construction rather than by four hand-kept lists agreeing.
 #
 # THE CLASSES LIVE HERE, NOT BESIDE THE LISTS, for the import graph: the
 # registry, the tuning table and the usage counter all read the specs, so the
@@ -353,8 +342,8 @@ class Surface:
     rendered by the tuning tool and the Settings UI. A floor cannot be moved
     sensibly by anyone, model or human, who does not know what the query is, what
     corpus it runs against, or how often it costs something. Those three facts
-    are exactly what separates these arms from each other, and they were
-    previously recoverable only by reading `plugin_context.py`.
+    are exactly what separates these arms from each other, and without them
+    here they are recoverable only by reading `plugin_context.py`.
     """
 
     name: str
@@ -375,13 +364,13 @@ class Surface:
     fires: str
     measured_model: str = "BAAI/bge-small-en-v1.5"
     measured_shape: int = 1
-    """What the SHIPPED defaults above were measured against (#4104).
+    """What the SHIPPED defaults above were measured against.
 
     A floor is a distance in one embedding model's geometry, over documents cut
     one particular way. Either can change, and when one does every number in
     this table describes something that no longer exists.
 
-    TWO FIELDS, NEVER ONE FUSED STRING (rule 149). A mismatch has to be able to
+    TWO FIELDS, NEVER ONE FUSED STRING. A mismatch has to be able to
     say WHICH half moved: a new embedding model and a re-cut document shape
     invalidate the same numbers for different reasons and call for different
     responses. `"<model>@<n>"` could only report that something changed, which
@@ -401,11 +390,11 @@ class Surface:
     budget_falls_back_to: str = ""
     """A budget key to inherit when this surface has none of its own set.
 
-    Only `write_path` uses it, and only because it USED to share auto-inject's
-    `top_k` outright. Giving it a key without this would silently reset the
-    budget of every install that had tuned the shared one — a behaviour change
-    delivered as a default, which is the shape of regression nobody reports
-    because nothing looks broken.
+    Only `write_path` uses it: it inherits auto-inject's `top_k` until its own
+    key is set. Without the fallback, an install that tuned the shared budget
+    would have it silently reset — a behaviour change delivered as a default,
+    which is the shape of regression nobody reports because nothing looks
+    broken.
     """
 
 
@@ -419,7 +408,7 @@ class Declared:
     """One line, for an agent reading a warning that names this source."""
 
     quiet_because: str = ""
-    """Set when silence over an active window is correct, saying why (#2475)."""
+    """Set when silence over an active window is correct, saying why."""
 
 
 @dataclass(frozen=True)
@@ -442,7 +431,7 @@ class RuleArm:
     """The telemetry `source`. MUST equal the registry and SURFACES key."""
 
     band: bool
-    """Keep only the hits within `_RULEHINT_BAND` of the top one (#3851)."""
+    """Keep only the hits within `_RULEHINT_BAND` of the top one."""
 
     compact_tail: bool
     """Lines after the first cite the rule instead of quoting its trigger."""
@@ -451,13 +440,13 @@ class RuleArm:
     """The top hit may hold the act for one read (`checkpoint_for`)."""
 
     preference_slot: bool
-    """Reserve one line for a preference that lost the ranking (#3894)."""
+    """Reserve one line for a preference that lost the ranking."""
 
     kind: str | None = None
     """Ask the ranker for one record kind only ("preference"), or every kind."""
 
     stop_only: bool = False
-    """The door shows nothing but the stop (milestone 458's reply moment).
+    """The door shows nothing but the stop (the reply moment).
 
     At the end of a turn there is no context to annotate: a hit either holds
     the reply for one read or reaches nobody. So only the checkpoint's rule is
@@ -471,9 +460,9 @@ class RuleArm:
     declared: Declared | None = None
 
 
-# Today's differences, reproduced exactly (milestone 456 step 2). Whether the
-# prompt arm should band, and whether the act arms should reserve a
-# preference, are step 7's questions — not settled by this refactor.
+# The arms differ on purpose where these flags differ. Whether the prompt arm
+# should band, and whether the act arms should reserve a preference, are open
+# questions to settle on telemetry, not by symmetry.
 PROMPT_RULE = RuleArm(
     "prompt_rule", band=False, compact_tail=False, checkpoint=False,
     preference_slot=True,
@@ -525,7 +514,7 @@ WRITE_PATH_RULE = RuleArm(
 REPLY_RULE = RuleArm(
     "reply_rule", band=False, compact_tail=False, checkpoint=True,
     preference_slot=False, stop_only=True,
-    # THE REPLY BACKSTOP (milestone 458, folded in from 456 step 8). Its floor
+    # THE REPLY BACKSTOP. Its floor
     # is a STOP bar, not a hint bar: at the end of a turn nothing can be shown
     # beside the reply, so a hit either holds the reply for one read or says
     # nothing. Hence a default at the checkpoint's level and a budget of one —
@@ -582,7 +571,7 @@ class RuleMoment:
     query: str
     project_id: int | None
     """The bound project, 0 or None when unbound. Logged as given, searched as
-    `project_id or None` — global rules plus that project's own (milestone 414)."""
+    `project_id or None` — global rules plus that project's own."""
 
     where: str = ""
     """How a line names the moment: "here", "to this Bash call", …"""
@@ -592,7 +581,7 @@ class RuleMoment:
     """Rules the session has been shown — rendered as repeats, not hidden."""
 
     held: frozenset[int] = frozenset()
-    """Rules the session has OPENED (#4100)."""
+    """Rules the session has OPENED."""
 
 
 @dataclass
@@ -624,7 +613,7 @@ def _telemetry_failed(source: str) -> None:
     The recorders are called DIRECTLY at each site, inside their own try,
     rather than through a wrapper taking the recorder as an argument: the
     registry test finds recorder call sites by their name, and a wrapper would
-    hide every site in this module from it (#3191's narrowing).
+    hide every site in this module from it.
     """
     logger.debug("retrieval telemetry failed for %s", source, exc_info=True)
 
@@ -637,7 +626,7 @@ async def _ranked(
 
     Returns (hits, kept, fresh): what the ranker returned, what the band kept,
     and the kept hits this session has not been shown. The row is written
-    HERE, before any caller can return early (#3497), with `results` the fresh
+    HERE, before any caller can return early, with `results` the fresh
     cut and `suppressed` counting both the band and the ledger, so a long
     session's all-repeat zero reads apart from a bar that turned everything
     away.
@@ -681,7 +670,7 @@ async def _ranked(
 async def _reserve_slot_for_preference(
     io: RuleIO, moment: RuleMoment, hits: list, *, floor: float,
 ) -> tuple[list, int | None]:
-    """Guarantee a preference one slot, if one clears the bar (#3894).
+    """Guarantee a preference one slot, if one clears the bar.
 
     THE ASYMMETRY THIS EXISTS FOR. A rule and a preference are not equally
     served by a shared score contest, because their losses are not equal:
@@ -697,33 +686,32 @@ async def _reserve_slot_for_preference(
     over the one whose loss is total, and it does so INVISIBLY: the rule that
     won is a legitimate hit, the telemetry looks healthy, and the only symptom
     is a preference that quietly never arrives. `reuse_slot` exists for the
-    same shape one corpus over (#2463), where snippets kept losing to project
-    records that merely resembled the query.
+    same shape one corpus over, where snippets lose to project records that
+    merely resemble the query.
 
     THE SLOT BUYS POSITION, NOT A LOWER BAR — same as `reuse_slot`, which also
     reserves at `cfg["threshold"]`. A weak preference cannot buy the slot, so
     silence stays the default and the reserved line is never worse than the
     ones it sits beside. If `preference_slot` later shows a stream of
-    near-misses, `best_available_id` (#3807) names which preference was
+    near-misses, `best_available_id` names which preference was
     refused and a separate bar becomes an argument with evidence behind it
     rather than a knob added on a guess.
 
     LEDGER REPEATS STILL COUNT AS REPRESENTED. A preference already on the
     session's ledger occupies the slot rather than being skipped for a fresh
-    one: it is still rendered (#3750), just with the tail that says so, and a
+    one: it is still rendered, just with the tail that says so, and a
     preference is the kind of record where being reminded is the point.
 
     Returns the possibly-extended hit list, and the id the slot spent — the
     caller needs that to keep each source's surfaced set matching its own log
-    row (#3668), since the slot logs under its own name.
+    row, since the slot logs under its own name.
     """
     if any(rule.kind == "preference" for _s, rule in hits):
         return hits, None
 
-    # ITS OWN SOURCE, and both sides of the trade logged. #2463's own finding
-    # is the warning rather than the precedent here: the hit that slot pushed
-    # OUT was in retrieval_logs while the query that pushed it out was not, so
-    # the slot could never be judged against what it displaced.
+    # ITS OWN SOURCE, and both sides of the trade logged. A slot whose query
+    # is not logged beside the hit it pushed out can never be judged against
+    # what it displaced.
     found, _kept, _fresh = await _ranked(
         io, moment, source=PREFERENCE_SLOT_SOURCE, floor=floor, limit=1,
         kind="preference",
@@ -748,11 +736,10 @@ async def _reserve_slot_for_preference(
     # ledger rather than taste. A displaced hit was RETURNED by the general
     # search and is sitting in that call's `retrieval_logs` row, but would not
     # have been shown — so `prompt_rule`'s surfaced set would stop matching
-    # its own log row, and #3668's identity would break for a reason nothing
-    # in the data explains. That identity is the cheapest true statement
-    # available about this pair of tables, and milestone #379 is what it costs
-    # to lose it: five steps planned against a gap that was two counters
-    # disagreeing, not a write path dropping rows.
+    # its own log row, and that identity would break for a reason nothing in
+    # the data explains. It is the cheapest true statement available about
+    # this pair of tables; without it, two counters disagreeing reads as a
+    # write path dropping rows.
     #
     # The price is one extra line, only when the general search already filled
     # the limit AND a preference cleared the bar without placing. Cheap, and
@@ -788,7 +775,7 @@ async def run_rule_arm(
             shown, _slot_id = await _reserve_slot_for_preference(
                 io, moment, shown, floor=floor,
             )
-        # `shown`, not `fresh` (#3750): a call whose only hit is a repeat
+        # `shown`, not `fresh`: a call whose only hit is a repeat
         # still has something to say, it just says it differently.
         if not shown:
             return RuleResult()
@@ -798,16 +785,16 @@ async def run_rule_arm(
                 rule, where=moment.where,
                 seen=rule.id in moment.exclude,
                 held=rule.id in moment.held,
-                # Rank decides volume (#3851): the ranker's best guess gets
+                # Rank decides volume: the ranker's best guess gets
                 # the trigger, the rest get cited.
                 compact=arm.compact_tail and idx > 0,
             )
             for idx, (_score, rule) in enumerate(shown)
         ]
-        # FRESH-ONLY (#3752): a repeat is a rendering decision, not a
-        # retrieval outcome, and counting it would inflate the denominator
+        # FRESH-ONLY: a repeat is a rendering decision, not a retrieval
+        # outcome, and counting it would inflate the denominator
         # pull_through is read from. It is also exactly what this call's own
-        # row recorded, so the two tables stay equal (#3668). RANKED, not
+        # row recorded, so the two tables stay equal. RANKED, not
         # ambient: the source is in `rule_usage.RANKED_SOURCES`.
         rule_ids = [rule.id for _score, rule in fresh]
         source = arm.source
@@ -838,20 +825,19 @@ async def run_rule_arm(
         return RuleResult()
 
 
-# ── The notes corpus (milestone 456 step 4) ──────────────────────────────
+# ── The notes corpus ─────────────────────────────────────────────────────
 #
 # The same stages over the other corpus: search, withhold what this response
 # already lists, split fresh from repeats, log the call before any early
 # return, band, reserved slots, surfacing rows. Two arms take them — the
 # prompt's menu (`auto_inject`) and the write path's match by meaning
-# (`write_path`) — and they differed only in which kinds they ask for, whether
+# (`write_path`) — and they differ only in which kinds they ask for, whether
 # a menu sits above them in the same response, and which slots they reserve.
 #
-# ONE DIFFERENCE FROM THE RULE ARMS IS KEPT, ON PURPOSE, FOR STEP 7. A notes
-# arm logs the cut BEFORE its band; a rule arm logs after it. #2085 chose the
-# first for notes (the log row is the candidate set a floor is tuned against,
-# the surfacing rows are what the reader saw) and #3851 the second for rules.
-# Both are recorded decisions, so this refactor reproduces both.
+# ONE DIFFERENCE FROM THE RULE ARMS IS KEPT, ON PURPOSE. A notes arm logs the
+# cut BEFORE its band; a rule arm logs after it. For notes the log row is the
+# candidate set a floor is tuned against and the surfacing rows are what the
+# reader saw; for rules the band is part of the cut being judged.
 #
 # What stays OUTSIDE: every lookup. The records named by number, the snippets
 # recorded at a path, the rulings, the design system and the shape ledger
@@ -860,7 +846,7 @@ async def run_rule_arm(
 
 # Margin gate: drop any hit more than this far below the top hit's score, so a
 # single strong match doesn't drag in a wall of barely-passing neighbours.
-# Twice the rule band, which is narrow because the rule corpus is flat (#3851).
+# Twice the rule band, which is narrow because the rule corpus is flat.
 _NOTE_BAND = 0.10
 
 
@@ -877,9 +863,8 @@ def _record_kind(note) -> str:
 
     A TASK ALSO CARRIES ITS STATUS, because for that kind alone the line is
     read as a claim about live work. A finished step and an open one rendered
-    identically is not a cosmetic gap: a done step was cited as a milestone's
-    open one on the strength of a line exactly like this, which carries an id,
-    a kind and a title and said nothing about where the work stood (#4154).
+    identically is not a cosmetic gap: a line carrying only an id, a kind and
+    a title invites a reader to cite a done step as open work.
     Only for tasks — a note or a snippet has no status to be wrong about.
     """
     if note.is_task:
@@ -895,22 +880,22 @@ def _record_kind(note) -> str:
     return note.note_type or "note"
 
 
-# WHAT A MENU LINE CARRIES (#4364): the record's NAME, its kind and System,
+# WHAT A MENU LINE CARRIES: the record's NAME, its kind and System,
 # and the WHOLE passage that matched. Metadata plus the evidence, rather than a
 # title asked to be both.
 #
 # The name, not the title. A snippet's or lesson's title is `name — when it
 # applies` by construction (`embeddings.trigger_title`), because that join is
 # what makes it rank on its situation. That is an EMBEDDING shape, and rendered
-# as a menu line it ran to 1,500+ characters — the trigger paragraph spent
+# as a menu line it runs to 1,500+ characters — the trigger paragraph spent
 # again on every line, and again on every repeat. The trigger still arrives
 # when it is what matched: it is in the chunk, and `_menu_passage` hands it
 # over when the title was the whole match.
 #
-# The whole passage, not 200 characters of it. The search already chose the
-# chunk that matched; the old cut kept its head and tail, and the head is the
-# title every chunk is prefixed with — so the reader got the title twice and
-# lost the middle, which is where the match was (lesson #4248). A chunk is at
+# The whole passage, not a head-and-tail cut of it. The search already chose
+# the chunk that matched, and its head is the title every chunk is prefixed
+# with — so a cut gives the reader the title twice and loses the middle,
+# which is where the match was. A chunk is at
 # most ~1.4 KB (`embeddings._CHUNK_CHAR_BUDGET`), and it is shown once: a
 # repeat is a one-line pointer (`_menu_seen_line`), not a second copy.
 
@@ -928,7 +913,7 @@ def _menu_name(title: str | None, note_type: str | None, data=None, body: str |
         from scribe.services.embeddings import untrigger_title
         from scribe.services.lessons import lesson_trigger
         trigger = lesson_trigger(SimpleNamespace(data=data, body=body or ""))
-        # One line even when the stored name is a story (#4797).
+        # One line even when the stored name is a story.
         return claim_line(untrigger_title(title, trigger).strip() or title)
     return title
 
@@ -972,16 +957,16 @@ def menu_entry(
 ) -> list[str]:
     """One record on a notes menu: its line, and what sits under it.
 
-    The prompt menu's two blocks — records named by number and records ranked
-    by meaning — used to write this out twice, differing only in what they
-    pass: a ranked line carries its score, a named one does not, and each puts
-    its own text under the line (the matched passage, or the record's opening).
+    Shared by the prompt menu's two blocks — records named by number and
+    records ranked by meaning — which differ only in what they pass: a ranked
+    line carries its score, a named one does not, and each puts its own text
+    under the line (the matched passage, or the record's opening).
 
-    A repeat is a POINTER, not a copy (#4364). The record is in this session's
+    A repeat is a POINTER, not a copy. The record is in this session's
     context already — the ledger is cleared at compaction, so "seen" stays true
-    — and re-rendering it spent its whole line again for nothing.
+    — and re-rendering it would spend its whole line again for nothing.
 
-    A superseded record is DEMOTED, not removed (#278), so one can still reach
+    A superseded record is DEMOTED, not removed, so one can still reach
     a menu, and when it does the reader has to be told: an agent handed stale
     material with nothing marking it acts on it with full confidence.
     """
@@ -1047,7 +1032,7 @@ class NoteArm:
     note_type: tuple[str, ...] | None = None
     task_kind: str | None = None
     include_global_kinds: bool = True
-    """Lessons are project-independent (#3730), so they join the candidate set
+    """Lessons are project-independent, so they join the candidate set
     from wherever they were learned."""
 
     scope: str = "browse"
@@ -1058,7 +1043,7 @@ class NoteArm:
     withholds: bool = False
     """A menu sits above this arm in the same response, and what it lists is
     kept out of the search (`exclude_ids`) — same-call duplication, which is a
-    different claim from the session ledger (#4101)."""
+    different claim from the session ledger."""
 
     tuning: Surface | None = None
     declared: Declared | None = None
@@ -1078,11 +1063,11 @@ AUTO_INJECT = NoteArm(
     ),
     declared=Declared("the notes menu offered at the prompt boundary"),
 )
-# Snippets AND recorded experience (#2246): an issue saying "we tried this and
-# it deadlocked" is prior art for the code about to be written. `task_kind`
+# Snippets AND recorded experience: an issue saying "we tried this and it
+# deadlocked" is prior art for the code about to be written. `task_kind`
 # keeps the open to-do list out — a task resembles the code and answers
-# nothing. Lessons too (milestone 385 step 5): the arm is kind-FILTERED, so a
-# kind absent here is unreachable, not merely outranked (#3702). No reserved
+# nothing. Lessons too: the arm is kind-FILTERED, so a kind absent here is
+# unreachable, not merely outranked. No reserved
 # slot: this arm fires before every Write and Edit, and the field is already
 # narrow enough that the 200:1 dilution a slot answers does not happen.
 WRITE_PATH = NoteArm(
@@ -1153,7 +1138,7 @@ class NoteMoment:
     """Searched and logged as `project_id or None`."""
 
     seen: frozenset[int] = frozenset()
-    """The session ledger: rendered as a pointer, never withheld (#4101)."""
+    """The session ledger: rendered as a pointer, never withheld."""
 
     named: frozenset[int] = frozenset()
     """Records this response shows by LOOKUP (named by number). Shown in their
@@ -1192,8 +1177,7 @@ def _note_band(hits: list) -> list:
 
     Computed over ALL hits, repeats included: the band measures distance from
     the top SCORE, and letting the ledger move that cutoff would make "you
-    were shown this" change what counts as relevant (#3851's axis
-    independence).
+    were shown this" change what counts as relevant.
     """
     if not hits:
         return []
@@ -1207,11 +1191,11 @@ async def _reserve_note_slot(
 ) -> tuple[list, int | None]:
     """Guarantee `slot.kinds` one line, if one clears the arm's own bar.
 
-    WHY A SLOT AT ALL (#2246, milestone 385 step 5). Ranking by raw cosine is
-    blind to what KIND of record answers what kind of ask, and the corpus
-    makes that fatal: Scribe's project records are about software work, so a
-    task about building a helper outranks the snippet that IS one. Snippets
-    were ~0.5% of the corpus when this was measured; no floor fixes 200:1. A
+    WHY A SLOT AT ALL. Ranking by raw cosine is blind to what KIND of record
+    answers what kind of ask, and the corpus makes that fatal: Scribe's
+    project records are about software work, so a task about building a
+    helper outranks the snippet that IS one. Snippets can be ~0.5% of a
+    corpus, and no floor fixes 200:1. A
     lesson crowded out is worse still — it exists only to be met at the moment
     it applies, so the arm that surfaces it IS its delivery.
 
@@ -1219,11 +1203,11 @@ async def _reserve_note_slot(
     floor and is NOT held to the band (the top score is the very thing these
     kinds lose to), so a weak record cannot buy the line.
 
-    ITS OWN SOURCE, from the first deploy. A guarantee has to be falsifiable,
-    and the hit a slot pushed out sits in the arm's row while the query that
-    pushed it out would otherwise be nowhere (#2463).
+    ITS OWN SOURCE. A guarantee has to be falsifiable, and the hit a slot
+    pushed out sits in the arm's row while the query that pushed it out would
+    otherwise be nowhere.
 
-    THE LEDGER IS NOT AN EXCLUSION HERE EITHER (#4101), but this call's own
+    THE LEDGER IS NOT AN EXCLUSION HERE EITHER, but this call's own
     menu is: a record already on it must not be shown twice, while one shown
     in an EARLIER call is exactly what a slot may spend itself on — a snippet
     relevant then and now is the reuse case, not a duplicate of it.
@@ -1270,7 +1254,7 @@ async def _reserve_note_slot(
         return menu, None
 
     slot_id = int(placed[0][1].id)
-    # FRESH ONLY, matching the row above: a ledger repeat is rendered (#4101)
+    # FRESH ONLY, matching the row above: a ledger repeat is rendered
     # but is not a new surfacing.
     if slot.books_own and slot_id not in shown:
         try:
@@ -1283,7 +1267,7 @@ async def _reserve_note_slot(
     if not slot.evicts:
         # IT EXTENDS, IT NEVER DISPLACES, siding with `preference_slot`: a
         # displaced hit sits in the arm's row, and evicting it would make the
-        # two tables disagree about the same call (#3668). And a record that
+        # two tables disagree about the same call. And a record that
         # does not bind should not throw a better-scoring one off the menu.
         return menu + placed, slot_id
     # Take the LAST line, never the first: the strongest overall hit is still
@@ -1316,15 +1300,15 @@ async def run_note_arm(
             kwargs["exclude_ids"] = set(moment.in_menu - scored)
         answered = await io.search(moment.user_id, moment.query, **kwargs)
         hits = [(s, n) for s, n in answered if int(n.id) not in moment.in_menu]
-        # WHAT THIS ARM WITHHELD AFTER THE SEARCH ANSWERED (#3739). The score
-        # the search reports is measured BEFORE that drop, so a record already
+        # WHAT THIS ARM WITHHELD AFTER THE SEARCH ANSWERED. The score the
+        # search reports is measured BEFORE that drop, so a record already
         # listed in this response could be logged as one the BAR turned away
-        # — live proof once read a "rejection" at 0.822 against a lowest
-        # acceptance of 0.6857. Whenever this removed anything, the honest
-        # best-available is null: "not measured on this call".
+        # — a "rejection" scoring well above the lowest acceptance. Whenever
+        # this removed anything, the honest best-available is null: "not
+        # measured on this call".
         withheld = len(answered) - len(hits)
         hits = hits[:budget]
-        # `results=fresh` and `suppressed` together (#3752): a rendered repeat
+        # `results=fresh` and `suppressed` together: a rendered repeat
         # is not a new surfacing, so it is COUNTED rather than reported, and an
         # all-repeat zero reads apart from a bar nothing cleared. A record
         # NAMED in this response is counted the same way — shown, but by the
@@ -1374,8 +1358,8 @@ async def run_note_arm(
         result.menu = menu
 
         # What SURVIVED the band — the menu the reader saw — where the call
-        # row holds the candidate set a floor is tuned against (#2085). FRESH
-        # ONLY, the row's own cut (#4101, #3668). A slot that books its own
+        # row holds the candidate set a floor is tuned against. FRESH ONLY,
+        # the row's own cut. A slot that books its own
         # line is not this arm's surfacing: counting it twice would leave the
         # slot's two tables describing different numbers of one event.
         own = {
@@ -1401,7 +1385,7 @@ async def run_note_arm(
         return NoteResult()
 
 
-# ── A rule reached through its lessons (milestone 440, #4633) ────────────
+# ── A rule reached through its lessons ───────────────────────────────────
 #
 # A rule's own document is written in the rule's words, which are general by
 # design; the situations that keep proving it are often closer to what a
@@ -1411,11 +1395,10 @@ async def run_note_arm(
 # It searches the NOTES corpus and answers with rules, which is why it sits
 # between the two halves of this module.
 #
-# ITS OWN SLOT, NOT A RULE SLOT, as a stated default. There was nothing to
-# measure until links are confirmed (#4632), so the asymmetry decides it: a
+# ITS OWN SLOT, NOT A RULE SLOT, as a stated default decided by asymmetry: a
 # via-lesson line that took a rule slot could push out a rule the ranker
 # matched DIRECTLY, a stronger claim displaced by a weaker one, while an extra
-# line costs one line. Logged as its own source so it can be judged (#4636).
+# line costs one line. Logged as its own source so it can be judged.
 
 VIA_LESSON_SOURCE = "rule_via_lesson"
 VIA_LESSON = RankedSource(
@@ -1493,7 +1476,8 @@ async def run_via_lesson_arm(
             chosen.append((s, rule, n))
             if len(chosen) >= VIA_LESSON_LIMIT:
                 break
-        # Logged whenever a search ran, results or not — the #3497 guard. The
+        # Logged whenever a search ran, results or not: a call that found
+        # nothing is the only evidence a bar is too high. The
         # score is the LESSON's, because the lesson is what was matched; the
         # best-available id is left off for the same reason, since the row's
         # results are rules and an id beside them would read as a rule id.
@@ -1536,7 +1520,7 @@ async def run_via_lesson_arm(
         return RuleResult()
 
 
-# ── The specs, read as lists (milestone 456 step 5) ──────────────────────
+# ── The specs, read as lists ─────────────────────────────────────────────
 #
 # What `retrieval_surfaces.SURFACES`, the ranked rows of
 # `retrieval_registry.POINTS` and `rule_usage.RANKED_SOURCES` are read from.
@@ -1558,20 +1542,20 @@ RULE_RANKED_SOURCES: tuple[str, ...] = (
 pull can confirm or refute it (`rule_usage.RANKED_SOURCES`)."""
 
 
-# ── The moment arm (milestone 458) ───────────────────────────────────────
+# ── The moment arm ───────────────────────────────────────────────────────
 #
 # A LOOKUP beside the ranked arms, not one of them. A rule mounted on a moment
 # arrives because somebody said it belongs there, not because the work's words
 # resemble it — which is the whole point: a rule about WHEN work is finished
 # never scores against what is said while finishing it. So there is no score,
-# no floor and no band, and it writes no retrieval_logs row (the rulings arm's
-# precedent, #4769): every score-shaped warning would misread a lookup. What it
+# no floor and no band, and it writes no retrieval_logs row (as the rulings
+# arm writes none): every score-shaped warning would misread a lookup. What it
 # records is the surfacing, with the moment beside each row, so a mount's
 # pull-through can be read per moment.
 #
 # It shares everything else with the ranked arms: the line renderer, the
-# session ledger (a repeat renders as a citation, never hidden — #3750) and
-# the fresh-only surfacing rows (#3752).
+# session ledger (a repeat renders as a citation, never hidden) and the
+# fresh-only surfacing rows.
 
 MOMENT_RULE_SOURCE = "moment_rule"
 
@@ -1608,7 +1592,7 @@ async def run_moment_arm(
     action that reached it — and every line says both ("at work.deliver,
     reached by `git push`"), so a moment that fired on the wrong act is
     visible in the line itself and can be corrected in the session with
-    `unmap_action` (step 2's ruling). Fresh rules come first and carry their
+    `unmap_action`. Fresh rules come first and carry their
     trigger; a rule the session was already shown is cited, not repeated in
     full. Fails open to an empty result.
     """