feat(adr): forbid embedded raw benchmark/spike data in ADR bodies (T7, #1925)

[FabioLeitao]

What changes

• ADR-0045 — new Decision item 9 (no embedded raw test/benchmark/spike data); Status history append-only line Amended for [P0][U1] adr(0045): proibir dado bruto de teste/benchmark no corpo de ADR — referenciar, nunca colar #1925.
• ADR-0000 — one sentence in the Sections table pointing to item 9.
• .cursor/rules/adr-trigger.mdc — bullet: link evidence, do not paste tables/logs/Operator decision (YYYY-MM-DD).
• Detector T7 — tests/adr_governance_support.py (embedded_experimental_data_violations).
• Tests — tests/test_adr_governance_phase1.py (T7 corpus, staged + lines, synthetic fixtures, auditor probe table).
• Plan — T7 row in docs/plans/PLAN_ADR_GOVERNANCE_ENFORCEMENT.md.
• docs/adr/INVENTORY.txt — regenerated via inv-adr.ps1 (body hash drift on ADR-0045).

What the detector flags

• A line Operator decision (YYYY-MM-DD) with optional Markdown prefixes (bullet, blockquote, bold).
• A Markdown table with three or more data rows whose cells look like measurement datasets (ms, µs/us, ×, x slower, req/s, MB/s).
• A fenced code block containing test-runner output (PASSED, FAILED, or pytest-style N passed / passed in Ns).

What it does not flag

• Prose numbers with a pinned artifact path or link (ADR-0078 style).
• status: failed inside a YAML sample block (no case-insensitive bare failed match).
• Option tables with % coverage splits or timeout tables with 5 s / 30 s (generic % / bare s are out of scope).

Transparency (required)

• The detector was run against all 92 real docs/adr/ADR-*.md files.
• ADR-0036 still contains a three-row microbenchmark table (µs/call) that the table heuristic would flag.
• The full-corpus test intentionally runs with check_tables=False, so legacy Accepted ADRs are not retroactively charged; the table rule applies only to added lines in new or amended ADRs (pre-commit staged diff).
• There is no grandfather.json.

Fixtures

• Synthetic good/bad excerpts under tests/fixtures/adr_t7_* — no text copied from private repos.

Ratification

• INVENTORY.txt was regenerated without -RatifiedBy; ADR-0045 SSHSIG ratification remains pending operator per ADR-0056.

Test plan

• ./scripts/check-all.sh exit 0

Closes #1925

---

Note

Low Risk
Documentation and pre-commit governance tests only; no runtime, auth, or data-path changes.

Overview
Adds ADR governance T7 (#1925): ADRs must record decisions in prose and link versioned evidence instead of embedding raw spikes, benchmark tables, or test-runner logs.

ADR-0045 gains Decision item 9 (forbidden shapes + allowed pinned paths) with a Status history amendment; ADR-0000 and .cursor/rules/adr-trigger.mdc tell agents to follow the same rule when materializing ADRs. embedded_experimental_data_violations in tests/adr_governance_support.py detects Operator decision (YYYY-MM-DD) lines, fenced pytest-style output, and (on staged + lines only) markdown tables with three or more measurement-like rows; the full-corpus check uses check_tables=False so legacy Accepted ADRs are not retroactively blocked. Phase 1 tests and synthetic good/bad fixtures exercise the detector; the enforcement plan and regenerated INVENTORY.txt reflect the ADR body edits.

^{Reviewed by Cursor Bugbot for commit 07e020b. Bugbot is set up for automated code reviews on this repo. Configure here.}

[FabioLeitao]

bugbot run

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 07e020b. Configure here.}

[cursor]

Staged table scan drops data rows

Medium Severity

_table_data_rows always discards the first two pipe-lines as header and separator. The staged T7 gate then runs that parser on concatenated + lines only, so adding or replacing three measurement rows on an existing table never meets the 3-row threshold and is not flagged.

Additional Locations (2)

• tests/adr_governance_support.py#L253-L274
• tests/test_adr_governance_phase1.py#L295-L307

 

^{Reviewed by Cursor Bugbot for commit 07e020b. Configure here.}

[cursor]

Unit regex matches English prefixes

Low Severity

BENCHMARK_TABLE_CELL_RE matches sec, ms, and us without a trailing boundary, so ordinary words such as seconds, messages, or users after a number count as measurement cells. A timeout or options table written in English can then fail the T7 gate.

 

^{Reviewed by Cursor Bugbot for commit 07e020b. Configure here.}

[FabioLeitao]

Corrigidos: inventário do ADR-0045 (5dd0d3a), tabelas parciais nas linhas adicionadas (646c4a3, bugbot média) e fronteira das unidades da regex (af58fae, bugbot baixa).