Review comments on #3603 landed after it merged: seven from CodeRabbit and four
from Codex. Two of the Codex findings are guarantees that were wrong in a way that
matters, and two are real defects in the fence scanners.

**The PII-safe log guarantee was overstated.** I documented
`error-safe-<date>.log` as recording only the error's runtime type, so it could be
shared without leaking user-authored content. `safeErrorDescription` renders
`'<message> (errorType=<Type>)'` — the caller's `message` goes in **verbatim**.
What the file actually omits is the raw exception string and the stack trace (whose
frames carry absolute paths and the system username). So it is shareable *because
callers are required to treat a log message as telemetry, never content* — the
contract `DomainLogger` states — not because the sink sanitises anything. Put a task
title in `message` and it is in the shareable log.

**`DerivedAgentState` is on the wake critical path after all.** I had trusted the
composite's own header — "Drives no production read; B6 flips reads onto this fold"
— but that sentence predates the cutover it predicted.
`AgentSyncService.reconciledAgentState` is called at the start of every task,
project, event, improver and day-agent wake; it folds markers and links over the
cached row so a value lost to last-writer-wins self-heals before the agent decides
anything. The concept now describes that, with the two boundaries that make it
precise — the merge is not a blind "log wins", and UI/service reads do stay on the
raw cache — and the stale header comment is corrected. That is the third time a
stale source comment seeded a wrong claim here, after `database_config_flags.dart`
and the six-catalog ARB list.

Two fence defects, both verified by fixture before and after:

- **A mixed closing run closed a block.** `closesFence` checked only the first
  character and the length, and the regex class allowed `` `~~ `` — so a malformed
  run closed a ``` block and the remainder of the file rendered as code while the
  check reported clean. CommonMark requires a uniform run.
- **Blockquoted fences were skipped in silence.** CommonMark strips container
  prefixes before recognising a fence, so `> ```mermaid` is a real diagram. Both
  scanners now strip blockquote markers, and the checker unwraps the body too —
  otherwise mermaid is handed `> > A --> B`. List-item containers are still not
  modelled, which is now stated rather than implied.

CodeRabbit's findings, all valid:

- `knowledge/log.md`'s single 2026-07-26 entry had accumulated ten bullets across
  the day, several recording claims that a later bullet in the same entry reversed —
  the privacy overshoot, "TaskStatus's five states", and expiry as merely
  "reported". Rewritten as the net outcome, and the convention now says a date's
  entry records that rather than the churn.
- The duplicated live/terminal split is **seven** files, not six.
- `verified` accepts either a mapping or a list; the doc named only `verified[].by`,
  which is the validator's message label, not the shape an author writes.
- `day_entries` was described as both "every persisted recording" and "newest 32".
  Both are true at different stages — `loadForDay` is unbounded, the section caps —
  and the two passages now say so.
- The Matrix zero-cache-limit lead-in still read as a present-tense assertion.

Not changed: Gemini's comment reports that its consumer review product has been
sunset, and codecov is unchanged at 99.15%.

…m round two

Both bots independently found the same defect in the fix from the previous commit,
and it is the worst kind: a **false green**.

**Container stripping ran on every line, including inside an open fence.** A
top-level fence whose body contained a literal `` > ``` `` had that marker stripped
and the line then read as the close, so an unclosed fence was reported as a cleanly
parsed block — verified by fixture before and after. CommonMark treats fence content
as literal; only the container a fence was opened *under* belongs to its body. Both
scanners now capture the opener's prefix and strip exactly that, and both suites
pin the literal-`>` case.

Four content findings, each checked against the code:

- **The logging diagram still labelled the safe log "runtime type only —
  shareable"** while the prose two paragraphs below said the guarantee is
  conditional. A reader trusting the diagram would still be told to share a file
  that can contain task titles. The node now states what the sink actually does:
  no raw error, no stack trace, message kept verbatim.
- **The reconcile is not history-independent.** I wrote that it "loads markers, not
  the full message log, so its cost does not grow with an agent's history".
  `getMessagesByKind(agentId, system)` passes no limit — `limit ?? -1` — and a
  milestone marker is appended on every completed wake, so the read and the fold
  grow with the number of wakes. Much smaller than the full log, still linear.
- **`knowledge/log.md`'s net-outcome entry still carried the superseded claim** that
  `DerivedAgentState` drives no production read — written before the same commit
  corrected it, which is precisely the contradiction that rewrite was supposed to
  remove.
- **The terminal-status inventory was wrong again**, and the shape was the problem:
  seven files became thirteen once the sealed-type tests were counted. Replaced with
  the five *idioms* and how to find them all, since a count in prose will keep
  drifting. Called out the quietest site — `_findResolution` returning null makes a
  wake run read as unresolved rather than erroring, so a new terminal status would
  silently distort the charts.

Tests: 96 Dart, 19 mermaid-checker.

… a fixed comment

Two findings on the previous commit, both mine.

**Recording the container as literal prefix text broke valid quoted fences.** The
marker's trailing space is optional in CommonMark, so `>flowchart TD` and `>``` `
under a `> ```mermaid` opener are the same container — but `startsWith('> ')`
matched neither, so the Dart validator reported a valid block as **unclosed** and
the JS scanner passed `>flowchart TD` to mermaid. Reproduced by fixture, then fixed
by tracking the opener's **depth** and stripping that many markers regardless of
spacing. Both suites now pin the varying-spacing case, and the literal-`>`-inside-a-
top-level-fence case still behaves — the two pull in opposite directions, so both
are asserted together.

**The projection concept sent readers to verify a sentence the same commit
deleted.** It said the composite's header "still says it drives no production read"
after that header had been corrected. Rewritten as history — the sentence outlived
the cutover it predicted, which is why it misled this concept twice — with
`agent_sync_service.dart` named as the authority for the read path.

Tests: 97 Dart, 20 mermaid-checker.