calibration: backfill phase 2 — GitHub-truth reversed labels (retro successor scan) + raw-context re-fetch

[JSONbored]

Problem

#8157's phase 1 (merged via its PR) backfilled 460 confidence-labeled cases from the historical ledger — but all confirmed: the D1 ledger provably contains zero derivable reversed labels (0 reversal events, 0 closes-later-merged), and no historical raw context. Two bounded GitHub-truth passes remain, deliberately split from phase 1 because both need the API:

1. Historical reversed labels via successor detection — run calibration: detect one-shot-culture reversals — bot-closed PR superseded by a merged successor (reversal_superseded) #8166's successor heuristics RETROSPECTIVELY: for each historical bot-closed PR, scan for a merged successor (same linked issue, or same author + file overlap). Each confirmed match retro-labels the original decision reversed (same provenance tagging as phase 1).
2. Raw-context re-fetch — for public-repo rows, re-fetch issue text / PR title/body / diff from GitHub to populate the corpus raw-context fields (bounded, resumable, rate-limit-aware). modelResponseText remains permanently out of scope.

Requirements

⚠️ Required pattern. Same pure-core/thin-IO split and idempotency discipline as phase 1 (scripts/backfill-calibration-corpus-core.ts): deterministic ids, INSERT OR IGNORE/targeted updates, backfilled + distinct provenance tags per pass, dry-run default with a report pasted here before any --apply.

• Depends on calibration: detect one-shot-culture reversals — bot-closed PR superseded by a merged successor (reversal_superseded) #8166's heuristics core landing first — this reuses that exact logic retrospectively, never a second similarity implementation.
• GitHub fetches must respect the shared budget (the rate-limit posture in gittensory-ratelimit conventions): resumable cursor, hard per-run cap.
• Conservative labeling only: a borderline successor match records nothing.

Deliverables

• Retro successor-scan pass (labels), gated on calibration: detect one-shot-culture reversals — bot-closed PR superseded by a merged successor (reversal_superseded) #8166
• Raw-context re-fetch pass (public repos only), separate flag
• Dry-run reports here before each apply

Links & Resources

#8082 (epic), #8157 (phase 1), #8166 (the heuristics this reuses), scripts/backfill-calibration-corpus-core.ts (the pattern)

Boundaries

No modelResponseText synthesis, ever. Nothing for private repos in the re-fetch pass. Applies remain maintainer-triggered.

maintainer-only — production-ledger data passes.

[linear-code]

JSO-5676

[JSONbored]

Pass A dry-run report (production, 2026-07-23)

Full scan of all 460 phase-1 backfilled close decisions (JSONbored/awesome-claude, closes 2026-06-11 → 06-22), successor listing depth budget-bound after the first run proved a fixed page cap silently truncates ("zero matches" ≠ "no successors").

scanned: 460   would-patch: 301   no-match/skipped: 159
match classes:
  same-PR reopened+merged   6   (definitive — the operator overrode the decision on its own target)
  same-author rework      167   (strong — the one-shot close of work its author then landed, ≥50% file overlap, ≤30d)
  shared-issue-only       128   (different author — routine duplicate competition, NOT reversal evidence)
GitHub requests used: 2671

Apply policy (the conservative bar this issue demands): apply same_pr_merged (6) + same_author (167) = 173 reversed labels; exclude the 128 shared-issue-only matches — in this culture the gate closing a losing duplicate while a different author's PR merges is the system working, and labeling those "reversed" would poison the corpus with ~28% false negatives. The CLI now enforces this default (--include-shared-issue-only is an explicit opt-in), and every applied row carries its class provenance (github_same_pr_merged / github_successor_scan) so config-era effects stay segmentable — several historical closes were faithful executions of since-changed rules, not judgment errors.

Applying to both stores (cloud D1 + the ORB server's Postgres, which carries the same seeded 460 targets) via a scan-once/apply-from-plan flow. Raw-context pass report follows separately.