fix(spac): make sec spac process reach a fixpoint

[sroussey]

loadGatedNoOpAccessions selects the filings whose known-SPAC-gated handler
recorded a success while writing nothing, so sec spac process can repair them.
For that to terminate, every branch has to be monotone: processing a filing
must write the artifact the branch keys on, and so take it out of the set. Three
branches were not, and each re-selected the same filings on every run forever.

fdb3e1e's commit message claims that cost is "self-limiting: they then carry
an artifact or a dead-letter entry, so the next run skips them." That claim is
false for all three.

1. 20-F — an FPI re-runs all six of its annual reports, forever

FORM_TO_EXTRACTOR_ID routes 20-F / 20-F/A to 25-15 so an FPI close
filing can record its combination. classifyListingRemoval answers "ignore"
for every ordinary annual 20-F, and (since b45ad0d) for every post-completion
one; processDeregistration then returns without writing. The 25-15 branch
selected any accession with no event, so a de-SPAC'd foreign private issuer with
six annual 20-Fs replays all six on every run, and gains one more each year.
backfillDescriptors.ts already got this right — the two had drifted.

A second, smaller instance in the same branch: processDeregistration returns
early on if (!args.filing_date), and a filing with no date was still selected.

2. 5.03-only 8-K

5.03 is in GATED_8K_ITEM_CODES via MILESTONE_ITEM_CODES, but
mapItemCodesToSpacEvents maps it to an event only when extractNameChange
finds a new registrant name in the narrative — and the narrative is non-null
only for a fetch escalated to the full submission, which
ProcessAccessionDocFormTask does for the redemption / LOI trigger items alone.
A 5.03-only 8-K therefore has no narrative by construction: it writes no event
on any run, and having no fullSubmissionText it runs neither detector, so it
leaves no dead letter either. Nothing it can produce would ever deselect it.

3. merger-proxy on the optional forms — the highest-volume one

The branch selected any merger-proxy accession with no spac_merger_extraction
row. But processMergerProxy deliberately writes nothing for a
MERGER_PROXY_OPTIONAL_FORMS filing whose merger section is absent
(skipMergerSection) — no row, no event, no dead letter, by design, because
most of those filings are annual and extension votes and dead-lettering each one
would bury the genuine failures. A plain DEF 14A reaches 575 distinct
SPACs, so every ordinary annual/extension proxy of every one of them was
re-selected on every run. mergerProxyDescriptor.filterTodo carried the
identical predicate and the identical non-convergence.

Also fixed

Repairing a gated 8-K took two runs. ProcessSpacTimelineTask computed the
gated set before the replay, and that set is empty for a CIK with no spac
row — which is exactly the state this command exists to repair. The S-1 that
mints the row is normally on the same timeline and is replayed moments later, so
every gated 8-K had already been filtered out of the one run that could fix it. A
CIK with 58 gated 8-Ks reported skipped: 58 and an empty timeline, and only a
second, identical invocation repaired it — with nothing saying so. The set is now
recomputed after the replay, minus everything already sent to the processor, and
whatever remains gets one capped serial pass in timeline order. One pass is
provably enough for the row-minted-mid-run case once the three fixes above make
the set convergent; the cap means it cannot spin if a fourth non-convergent shape
is ever introduced.

Quadratic dead-letter scan. loadDetectorAccessions called
listAll("redemption") and listAll("loi") — every row of each extractor across
every CIK, materialized, memoized only within one loadGatedNoOpAccessions call,
i.e. per CIK. A batch over ~1,500 SPACs loaded ~60k rows per CIK. It now uses a
new ExtractionDeadLetterRepo.listByAccessions, built like the existing
listPendingByAccessions (chunked in-lists at MAX_IDS_PER_QUERY, no status
filter — a resolved row is the evidence), fed only the 8-K accessions that
survived the item-code and event tests.

Changes

• processDeregistration.ts — new exported listingRemovalNeedsWork, the single
  predicate both spac process and sec extractor backfill 25-15 select on, so
  the two cannot drift again.
• gatedNoOpAccessions.ts — 25-15 branch delegates to it; 5.03 dropped from
  the gated item codes (with the reason stated, since a future reader adding it
  back is the regression); merger-proxy branch gains the dead-letter conjunct;
  detector lookup scoped to the issuer's accessions.
• Form_DEFM14A.storage.ts — the skipMergerSection branch records a RESOLVED
  SECTION_NOT_FOUND entry, via a new ExtractionDeadLetterRepo.recordResolved
  (one put, attempts: 0, first_seen_at preserved). The entry never reaches
  sec extractor dead-letters, which lists pending entries, so the worklist an
  operator reads is untouched; the volume is comparable to the auto-resolved
  MODEL_EMPTY rows the redemption / LOI detectors already write per trigger 8-K.
• ProcessSpacTimelineTask.ts — replay loop factored into replayFilings, plus
  the capped repair pass. skipped is now counted against what was actually sent
  to the processor across both passes.

Invariant

Every predicate touched keys on an artifact the handler writes (event /
extraction row / dead-letter row) or on the classifier agreeing with what is
already recorded. Two "reaches a fixpoint: a second invocation processes nothing"-style tests hold that line; the seven new negative tests below each
fail on main.

This PR strictly reduces work and changes no data semantics: nothing is
selected that was not selected before, no extractor behaviour changes, and the
only new write is a resolved dead-letter row recording an outcome that was
previously left untraced.

Tests

New, all verified failing on main before the fix:

• gatedNoOpAccessions.test.ts — annual 20-F, post-completion 20-F, 5.03-only
  8-K, optional-form proxy with a resolved SECTION_NOT_FOUND, and a scoping
  test that spies on the dead-letter storage and rejects any unfiltered
  { extractor_id }-only query. Plus two positives pinning that the exclusions
  are narrow: a 5.03,8.01 8-K and a proxy with no evidence at all are still
  selected.
• ProcessSpacTimelineTask.test.ts — gated 8-K processed in the same run as the
  S-1 that mints the row; a second invocation processes nothing; no second pass
  under --dry-run.
• Form_DEFM14A.storage.e2e.test.ts — the extension-proxy case now asserts the
  resolved entry exists and that listPending("merger-proxy") does not contain it.
• backfillDescriptors.test.ts — general proxy with a legitimately absent merger
  section is not re-selected; a 20-F the classifier ignores is skipped while the
  Form 15 beside it is still selected.

$ bunx vitest run src/task/spac/ src/task/forms/backfillDescriptors.test.ts \
    src/sec/forms/exchange-listing-withdrawal/
 Test Files  16 passed (16)
      Tests  218 passed (218)

$ bun run format-check
All matched files use Prettier code style!

$ npx tsc --noEmit     # clean

Wider runs: src/task/ 481 passed / 15 skipped, src/storage/ +
src/sec/forms/miscellaneous-filings/ 969 passed. src/sec/ has one failure,
parseOfferingTables.corpus.test.ts > loads committed S-1 fixtures — a 30s test
timeout, reproduced on a clean main checkout and unrelated to this change.

---

🤖 Generated with Claude Code

https://claude.ai/code/session_01LowBJQsCghLDiHwPN6FgUT

---

Generated by Claude Code