python-standards content backlog: candidate rules from dogfooding + writing-skills compliance

[vircung]

python-standards content backlog — the python mirror of #6. Sources: the
2026-07-23 dogfooding and writing-skills review, and the 2026-07-27
review-contract gate dogfooding (three python projects, five runs). None
of these is a defect in 0.2.0 — the plugin produces correct,
contract-conforming reviews; these are coverage and authoring gaps.

P1 — candidate rules (strongest first, by multi-project recurrence)

• Dead code (modules, functions, dead params/branches) — recurred
  in every audited project; style-no-dead-code / layout-dead-code.
• Atomic state writes / shared-state thread safety — non-atomic
  writes of resume/recovery state files (critical-level evidence) and
  shared caches mutated under a thread pool; e.g.
  layout-shared-state-thread-safety + an atomic-file-writes rule.
• cli-subprocess-timeout — with per-path sub-rules: bulk path →
  important, one-shot → minor (one run reported the same class twice
  with split severities and flagged the split itself as the rule
  author's main insight).
• standard-retry-classification (critical) — all errors treated
  as permanent, no transient-vs-permanent classification before a
  skip-list write.
• style-exception-fidelity (important) — except Exception masking
  a precise exception; style-metric-correctness (important) — metric
  not measuring what it declares; style-query-safety-guard
  (important) — builder silently returning an unbounded query.
• layout-destination-collision-safety (critical) — silent overwrite
  of already-ingested data on a name collision;
  layout-identity-key-strength (important).
• Minors: layout-single-version-source,
  cli-no-source-string-interpolation, cli-retry-diagnostics,
  cli-one-row-per-item, test-no-source-constant-copies,
  test-helpers-not-in-test-modules, style-comments-state-constraints,
  typing-private-helper-clarity, pool-resource return without
  try/finally, duplicate logging of one event, one-state-two-locks.
• Consider a new python-data-integrity skill — one run proposed
  it for the four critical-level data-integrity gaps (atomic writes,
  retry classification, collision safety, identity keys) instead of
  scattering them across style/layout.

Writing-skills compliance (python part)

• Description keyword gaps: python-code-style + docstring;
  python-testing + mock/patch; python-typing + Any / type: ignore /
  mypy; python-project-layout + mise.
• python-plan-review description rewritten to the "Use when…" form.
• Citation-boilerplate decision (tersen vs centralize) — family-wide,
  decided once with salesforce-standards content backlog: missing rules, deliberate sub-rules, writing-skills compliance #6.
• Judgment-word predicate → observable predicate: "obvious from one
  screen of context" (python-typing).

Shared family decisions (made once, applied here and in #6)

• README source-bridge sentence: per-rule provenance lives in each
  rule tag's source: field.
• Audit-mode semantics for "new-*"-scoped rules (salesforce-standards content backlog: missing rules, deliberate sub-rules, writing-skills compliance #6 P4).

New rules follow the #8 tag convention and its authoring rubric;
report-contract behavior is owned by the texts shipped in #10.