docs(launchpad): stop quoting the suppression trigger phrase verbatim

[serina-mcfall]

Summary

launchpad/plans/2026-08-12-issue-117-review-dimensions.md quotes the literal suppression-attack trigger phrase verbatim, which trips detect.py's _SUPPRESS regex and fails check_step6.py's false-positive control — the reason controls CI is red on #241 and #242, both of which inherit this file unchanged from launchpad. This rewords the one offending paragraph to describe the verification abstractly, the way CONTAINMENT.md already does, without changing its substantive claim.

Related issue

Closes #244

Issue type

Bug

---

Agent provenance

Field	Value
Harness / provider	Claude Code
Model	claude-sonnet-5
Session reference	N/A - harness does not expose a session URL
Initiating human	@serina-mcfall

Objective

Make launchpad/plans/2026-08-12-issue-117-review-dimensions.md pass check_step6.py's zero-false-positive control, unblocking the controls CI check inherited by #241 and #242.

Impacted components

launchpad/plans/2026-08-12-issue-117-review-dimensions.md

Approach and rejected alternatives

Reworded the offending paragraph to state the verification by the sentence's grammatical shape (matching CONTAINMENT.md's own established phrasing for this exact rule), instead of quoting the literal trigger phrase. Rejected: weakening detect.py's _SUPPRESS regex or excluding this file from check_step6.py's corpus scan — both would hide a real detector capability from the control that exists specifically to prove it, and the docstrings in detect.py/CONTAINMENT.md already establish that the correct fix for this exact situation is rewording the prose, not the detector or the test.

Verification

Command run:

cd launchpad/review-agent
python3 -c "
from detect import detect
text = open('../plans/2026-08-12-issue-117-review-dimensions.md', encoding='utf-8').read()
findings = detect(text, 'pr_body')
print(len(findings), 'finding(s)')
"
python3 check_step6.py
python3 run_controls.py

Raw output:

0 finding(s)

PASS  zero false positives on ten upstream records (got [])
PASS  review-heavy corpus assembled (15 documents)
PASS  zero false positives on this repo's review-heavy text (got [])

  recall: 28/35 caught ['approve', 'delimiter-breakout', 'skip-review', 'suppress-a-finding']
          7/35 missed ['paraphrase'] — model layer's responsibility (#117)

PASS  the detector catches something
PASS  IN-SCOPE held-out payload is detected — the detector generalises, not memorises
PASS  in-scope held-out findings are Blocker
PASS  OUT-OF-SCOPE held-out payload is missed, exactly as CONTAINMENT.md documents
PASS  both held-out payloads are genuinely absent from the attack corpus
PASS  every finding carries severity Blocker
PASS  every finding names its own entry point
PASS  CONTAINMENT.md documents the limit
PASS  CONTAINMENT.md states the real counts (28 caught / 7 missed)
PASS  contain.py --json carries injection_attempt (got {'injection_attempt'})
PASS  every finding in CLI output is Blocker

0 failure(s)

PASS  check_contract.py
PASS  check_step2.py
PASS  check_step3.py
PASS  check_step45.py
PASS  check_step6.py
PASS  suite.py
PASS  check_step8.py
PASS  check_step9.py
PASS  check_step11.py
PASS  check_invariants.py
PASS  check_mutations.py
PASS  check_flag_guard.py
PASS  check_fetch_states.py

13 passed, 0 failed, 0 skipped

Independently reproduced by Codex (gpt-5.6-sol), which re-ran detect() against both the original commit 4a352e36d (1 finding, matching the reported CI evidence) and the fixed worktree (0 findings), then ran check_step6.py directly (its own sandbox lacked gh auth for the aggregate run_controls.py, so it ran the network-gated control standalone instead and confirmed 0 failures). Its explicit verdict:

VERDICT: stands — the change minimally removes a documented corpus-convention violation while preserving detector behavior, test coverage, and the plan's load-bearing contract.

• Tests or checks were run and the raw output is pasted above
• The diff is confined to the scope of the linked issue
• No secrets, keys, tokens or hostnames were added to tracked files

Not verified

Did not confirm that #241 and #242 turn green after merge — that requires those branches to rebase/merge onto launchpad after this lands, which is out of scope for this PR. Did not run the full desktop/mobile/relay CI matrix locally (unaffected — this change touches only one markdown file under launchpad/).

Security implications

None. This is a documentation-only change to a planning doc's prose; it does not touch detect.py's matching logic, contain.py's escaping/rendering, or any code path that processes untrusted PR content.

Escalations

None — the fix is a same-convention rewording, not a design decision, and the cross-vendor review found no open questions.

[tucktuck101]

@serina-mcfall, this specific one has conflicts, please investigate and address 🙏

[serina-mcfall]

Superseded — PR #245 fixed the identical problem (same file, same paragraph, different wording) and merged first as commit b03322f. check_step6.py already passes cleanly on current launchpad (confirmed: 0 failures). This PR's conflict is that duplication, not something to merge-resolve. Closing without merging; #244 (which this closes) is a duplicate of #246, which #245 already closed.