[dekobon]

Summary

Two passes over every durable Markdown doc (root and crate READMEs, STABILITY.md, RELEASING.md, CONTRIBUTING.md, SECURITY.md, docs/, and the book source), each reviewed by an independent fresh-context agent. One commit; 76 files.

Accuracy

Roughly 165 claims no longer matched the code. Each fix is recorded with its evidence (oracle file:line or command output) in the sweep's findings; the ones that change what a reader would do:

• find / count take node types through -t/--type, not positionally; the CLI README was missing vcs, init, diff, diff-baseline and exemptions.
• -t 0 does not match ERROR (that is 65535); check baselines never filter report hotspot tables; the cognitive rule charged try where only catch pays; CC was presented as the minimum branch-coverage test count.
• The documented cargo mutants alias is unresolvable under cargo 1.99 (fix(tooling): the cargo mutants alias in .cargo/config.toml is unresolvable under cargo 1.99 #1710); the doc now calls cargo-mutants directly.
• The Python pages gave detection precedence as extension, shebang, mode line (it is extension, mode line, shebang); InvalidDiffError was attributed to vcs.commit; the "zip works" example did not.
• partial_risk_score is not "always lower" than a commit score (jit.rs says the opposite); the REST route list omitted HEAD; /metrics omitted npm/npa.
• RELEASING.md still described the repository as pre-public with the publish variables unset, named two jobs that do not exist, and a [workspace.dependencies] pin that does not.
• STABILITY.md listed a Metrics type removed at 2.0 and the pre-2.0 Span keys. It now also names both non-MSRV (breaking) entries in the 2.x line (fix(wire): recursive Serialize/Drop on FuncSpace lack the #700 de-recursion #1056, fix(py): batch docs promise 1:1 results under skip_generated=False, but tiny/binary files still drop slots #1238).
• Hardcoded test and corpus counts become approximate language or a presence matrix, per docs/conventions/documentation.md. Historical records (threshold_corpus.md, the output-normalization postmortem, measured figures in benchmarking.md and fuzzing.md) keep every number, date and revision.

Readability

Under docs/conventions/documentation.md: prose em dashes removed (quoted CLI output and marker syntax keep theirs), first-mention links for external tools and standards, orientation openings where a page started cold, and check.md, RELEASING.md, benchmarking.md and fuzzing.md reordered so the common question comes first. The convention now names the em-dash replacements.

Pinned names

Every renamed heading keeps its old slug through an explicit {#id}. Unchanged: the ## headings doc_anchors_resolve_to_book_headings pins, the threshold tables doc_summary_tables_match_the_default_table pins, every {{#include}} range, and the version pins check-versions gates. A before/after audit of every fragment link in the tree is identical.

Also

• The unused FreeBSD packaging templates go with the job ci: re-add x86_64-unknown-freebsd binary target via vmactions/freebsd-vm #346 removed.
• README.ja.md mirrors the factual corrections and new links; heading order was unchanged.
• README.md's relative links still resolve only on GitHub (crates.io, the deb/rpm and the tarball ship it too); left as is.
• Out-of-scope drift the sweep surfaced is filed as fix(cli): help text names deprecated flags and pre-#597 option layout (ships in man/) #1709, fix(tooling): the cargo mutants alias in .cargo/config.toml is unresolvable under cargo 1.99 #1710, chore(tooling): stale literals in worktree-setup, grammar scripts and workflow comments #1711, chore: stale doc comments name removed items and pre-2.0 behaviour #1712, docs(skills): add-lang skill is behind the live mk_langs! shape, metric count and role tables #1713.

Verification

• make pre-commit: BCA_GATE: pass (gate=pre-commit).
• rumdl check clean on all 80 tracked Markdown files; make book builds; make check-versions passes; default_thresholds (9) and doc_anchors (1) tests pass.
• Review: eight accuracy agents, one per doc group; astra-review verified every changed hunk (112/123 clean, 8 findings applied); two sol-review samples of 125 unchanged claims (13% miss rate, all misses fixed); one astra-review per readability group, findings applied. Review-flagged "restore HEAD" items were rejected where HEAD predated the accuracy pass.