feat(launchpad): adjudication fixtures, four of five genuinely produced (#118 STEP 8)

[serina-mcfall]

Related issue

Refs #118

Issue type

Task

#118 STEP 8 — adjudication fixtures, four of five genuinely produced

Implements STEP 8 of launchpad/plans/2026-08-13-issue-118-adjudication.md, stacked on the not-yet-merged STEP 7 branch (#267). Adds launchpad/review-agent/fixtures/adjudication/ (a generator, four fixture documents, a crafted-surfaces payload, and PROVENANCE.md), plus test_adjudication_fixtures.py — and corrects STEP 8's own text in the plan.

STEP 8's premise was dead, and correcting it was part of the work

The step's text said:

"THESE ARE SYNTHESISED, NOT RECORDED, AND THAT IS A KNOWN WEAKNESS. #117's producer does not exist — there is no run_dimensions.py anywhere... Until then the suite's coverage of real producer output is zero."

None of that has been true since #117 merged. The plan's own BUDGET section already anticipated this and instructed that "if their text still describes document-only synthesis, that text is what needs updating, not this correction." So it is corrected here — struck through rather than deleted, matching the convention used throughout the plan's ALREADY TRUE and BUDGET sections.

What is real, measured before building

fixture	isolates	provenance
line-anchored-findings.json	three reports, one finding per dimension, all anchor line and the dedupe case	real — replays recordings/paraphrase/*
pr-anchored-finding.json	a pr-anchored finding (file/line null) alongside a line-anchored one	real — replays recordings/claim-vs-evidence/*
mixed-report-statuses.json	one failed, one clean, one with findings	real — replays recordings/secrets-and-access/*, failure raised through the real handler
containment-all-kinds.json	all three containment kinds, seven-key states, zero dimension findings	crafted surfaces, real pipeline

Two things worth drawing out:

paraphrase is a natural dedupe fixture. All three dimensions independently reported the same defect at crates/buzz-relay/src/gate.rs:42, with three different finding_ids (dimension is a hash input). That is STEP 7's dedupe case occurring in real recorded output rather than anything hand-built — so one document legitimately isolates two of STEP 8's named behaviours. Committing it twice under two names, or inventing a second document, were the dishonest alternatives.

The containment fixture is the one genuine exception, and the split is stated rather than blurred. No pre-existing fixture trips the detectors — all eight were checked (benign.json, captured-pr.json, payloads.json, and all five under fixtures/dimensions/), every one yielding zero containment findings. So its surfaces are crafted; its containment block and states map are then produced by the real contain.render. Crafted input, real pipeline — never described as recorded.

The claim is checkable, not asserted

Regenerating reproduces the committed bytes exactly. Every nonce derives from the relevant recording's own _provenance.seed via contain.make_nonce(seed=...), never freshly randomised. That is what turns "these came from real recorded output" into something a reviewer can falsify rather than take on trust.

Verification

• 18 new tests; run_controls.py 13/13; all eight suites green
• Each fixture accepted by findings.validate with zero violations, and by run_adjudication.py at exit 0 with output passing both verdicts.validate and findings.validate
• Containment fixture: all three kinds, exactly seven states keys matching contain.ENTRY_POINTS

Reviewed (serina:review-code, independent pass) — it attacked the central claim rather than accepting it: traced individual findings byte-for-byte from their recording files into the committed fixtures, confirmed the failed report's error.reason matches _collect_report's real exception branch, independently re-ran the eight-fixture containment check, and verified byte-reproducibility out-of-band with its own diff. No findings.

One item it raised as context rather than a finding is fixed in 888001977: the generator cited "Serina's call recorded in this repo's session notes" — a source no reader can open. Replaced with the argument made from what the recording actually contains. This plan carries a standing rule that evidence is what git show returns from a named commit, adopted after three separate rounds of cross-issue citation failures; a pointer to out-of-repo notes is weaker than all three.

Draft, per this repo's pr-gate hook — review-final runs at STEP 12, end of the chain.

🤖 Drafted by Claude Code for @serina-mcfall.