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

The silent-outage story (#2663), the ranked/ambient split (#2477), the
transfer evidence (#3735) and the attach_usage consolidation (#4230) are
stated as the property each guards, without the incident.

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

src/scribe/services/note_usage.py

@@ -15,13 +15,12 @@ Design notes (mirrors retrieval_telemetry, for the same reasons):
   - Writes are fire-and-forget. `record_surfaced` / `record_pulled` extract
     plain ints synchronously and schedule the insert as a background task, so
     telemetry never adds latency to — or can break — the surface it observes.
-  - Failures degrade, but they must not degrade SILENTLY. The original version
-    swallowed everything into logger.debug, and the deployed instance ran with
-    every counter at zero for weeks while surfacing demonstrably fired — a
-    total outage indistinguishable from "nobody uses this" (#2663). A
-    subsystem whose every failure mode is invisible cannot report its own
-    death, so failures now log at WARNING and drop one AppLog error row per
-    process per site, where the admin UI shows it.
+  - Failures degrade, but they must not degrade SILENTLY. Swallowed into a
+    debug log, a failure leaves every counter at zero while surfacing fires —
+    a total outage indistinguishable from "nobody uses this". A subsystem
+    whose every failure mode is invisible cannot report its own death, so
+    failures log at WARNING and drop one AppLog error row per process per
+    site, where the admin UI shows it.
   - Reads (`usage_for_notes`) are NOT fire-and-forget — a readout the caller
     awaits, aggregated in one round-trip for a whole page of snippets rather
     than per row.
@@ -151,7 +150,7 @@ def record_pulled(
     `project_id` is where the READER was — the caller's active project, never
     the record's own. Compared against the record's `project_id`, it answers
     whether this was opened somewhere other than where it was written, which
-    is the evidence that a record TRANSFERRED (milestone 385, #3735).
+    is the evidence that a record TRANSFERRED.
 
     Unlike a surfacing arm, a getter only knows what it was handed, so this
     stays optional and null is an ordinary answer meaning "not reported". A
@@ -178,7 +177,7 @@ def record_pulled(
 # top-N-by-recency happen to be; the skill sync installs every Process the
 # operator can reach. Counting those alongside auto-inject would make a note's
 # surfaced_count dominated by "it was recently updated in a project you
-# opened", and dead-weight detection would read that as popularity (#2477).
+# opened", and dead-weight detection would read that as popularity.
 # They still matter — a pull that follows one must not float unattributed — so
 # they land in their own bucket rather than not landing at all.
 AMBIENT_SOURCES = ("enter_project", "process_skill_sync")
@@ -191,14 +190,13 @@ def empty_usage() -> dict:
     reads as "never surfaced, never pulled" rather than as a missing key.
 
     `surfaced_count` is RANKED surfacings only — a scored surface chose this
-    record. `ambient_count` is the rest (see AMBIENT_SOURCES). The split is the
-    readout half of #2477: the "high surfaced, zero pulls → dead weight"
-    reading is only valid over surfacings that were choices.
+    record. `ambient_count` is the rest (see AMBIENT_SOURCES). The split exists
+    because the "high surfaced, zero pulls → dead weight" reading is only valid over surfacings that were choices.
 
     `surfaced_away_count` / `pulled_away_count` are the subsets that happened
     on a project other than the one the record was written on — the evidence
-    that a record TRANSFERRED, which is the claim the lesson kind rests on
-    (milestone 385, #3735). Counted only where both projects are known: an
+    that a record TRANSFERRED, which is the claim the lesson kind rests on.
+    Counted only where both projects are known: an
     event with no reader project, or a record with no project of its own,
     cannot speak to "away" and is left out rather than guessed.
     """
@@ -231,9 +229,8 @@ async def usage_for_notes(note_ids: list[int]) -> dict[int, dict]:
     # distinct source. ONE labelled expression, grouped by its label — a second
     # case() instance in GROUP BY renders with its own expanding-IN bind names
     # under asyncpg, so the database sees two DIFFERENT expressions and rejects
-    # the query with a GroupingError. That rejection was swallowed, which is
-    # how every counter read zero in production while the writes were landing
-    # fine (#2663).
+    # the query with a GroupingError — and a swallowed rejection reads as
+    # every counter at zero while the writes land fine.
     ambient = case(
         (NoteUsageEvent.source.in_(AMBIENT_SOURCES), True),
         else_=False,
@@ -262,7 +259,7 @@ async def usage_for_notes(note_ids: list[int]) -> dict[int, dict]:
             # it needs the join to `notes` and the first must keep counting
             # events whose note has since been deleted (the table is FK-free
             # so that evidence outlives the row). Ranked surfacings only, for
-            # the same reason surfaced_count is (#2477).
+            # the same reason surfaced_count is.
             away_rows = (
                 await session.execute(
                     select(
@@ -284,7 +281,7 @@ async def usage_for_notes(note_ids: list[int]) -> dict[int, dict]:
     except Exception:
         # A telemetry readout must not be able to break the list it decorates —
         # but it must say it failed, or a broken readout is indistinguishable
-        # from a corpus nobody uses (#2663).
+        # from a corpus nobody uses.
         await _report_failure("readout")
         return out
 
@@ -335,15 +332,12 @@ def _row_id(row: dict, key: str) -> int | None:
 
 
 async def attach_usage(rows: Sequence[dict], *, key: str = "id") -> None:
-    """Add `usage` to every row of a payload a door is about to return (#4230).
+    """Add `usage` to every row of a payload a door is about to return.
 
-    The one seam both doors and every record kind share. Before this, four call
-    sites carried their own copy of the same lines — two list routes and two
-    detail routes — and `/api/knowledge`, which is the list a person ACTUALLY
-    browses notes and lessons in, was about to become a fifth. That is how the
-    chip came to reach two record kinds out of four while a service named
-    `usage_for_notes` worked on all of them: each door read fine on its own,
-    and nobody was comparing them.
+    The one seam both doors and every record kind share — list routes, detail
+    routes and `/api/knowledge`, the list a person actually browses notes and
+    lessons in. A copy per call site lets the chip reach some record kinds and
+    not others while each door reads fine on its own.
 
     ONE AGGREGATE FOR THE WHOLE PAGE. `usage_for_notes` is a single GROUP BY
     over the id set; calling it per row would be N+1 by construction, which is
@@ -360,11 +354,10 @@ async def attach_usage(rows: Sequence[dict], *, key: str = "id") -> None:
     reports it through `_report_failure("readout")` and returns the zero-filled
     map, so a broken readout degrades without breaking the list it decorates.
     Wrapping it again would swallow the REPORT along with the error, and a
-    silently-swallowed readout failure is exactly #2663 — every counter reading
-    zero in production for weeks while the writes landed fine.
+    silently-swallowed readout failure reads as every counter at zero while
+    the writes land fine.
 
-    Mutates in place and returns None, matching how the call sites already used
-    it: these rows are the payload, not a copy of it.
+    Mutates in place and returns None: these rows are the payload, not a copy of it.
 
     A detail payload is just a one-row list — `await attach_usage([data])` —
     so the single-record doors share this seam rather than keeping a second