D8: restore real type checking across the reactor↔afi-core boundary

[Gio2050]

Executes D8 from reports/WORKLIST.md — the highest-leverage open item.

⚠️ Merge AFI-Protocol/afi-core#28 FIRST. This PR's CI checks out afi-core at its default branch with no ref:, and the new strict typecheck step fails against afi-core@main until the indicators widening lands. A red CI here before that is expected, not a defect.

The problem was subtler than "resolution is broken"

typings.d.ts declared 16 ambient modules, 7 of which typed the entire afi-core boundary as any, alongside strict: false and tests excluded from tsconfig.

The hypothesis was that moduleResolution: "node" couldn't read afi-core's exports map. That was half right. From tsc --traceResolution:

Specifier	Resolved under "node"?	Shim?
analysts/froggy.enrichment_adapter.js	✅ → real afi-core source	yes
analysts/froggy.trend_pullback_v1.js	✅ → real afi-core source	yes
validators/UwrProfileLoader.js	✅ → real afi-core source	yes
validators/UniversalWeightingRule.js	✅ → real afi-core source	no
afi-core/analyst	❌	yes
afi-core/decay	❌	yes

Only two shims were load-bearing. An ambient declare module beats successful module resolution — four specifiers already resolved to real afi-core source and were still bound to any. Switching resolution alone changed nothing (0 errors); the shims had to be deleted.

Two consequences nobody could see: the reactor was compiling afi-core's source inside its own program under strict: false (laxer than afi-core's own strict: true build), and three shims declared modules the reactor never imports at all.

Results — 45 errors: 2 REAL BUG · 13 DORMANT · 30 NOISE

Removing shims reduced the count 45 → 21 while increasing type coverage — the clearest proof they were never load-bearing. 6 of 7 third-party shims shadowed types the package already ships; only input genuinely has none.

Change	Errors
baseline (16 shims, strict:false, src only)	0 ← the problem
bundler resolution, shims still present	0
+ 7 afi-core shims deleted	3
+ noImplicitAny → strict → tests included	45
− ajv/ajv-formats shims (real ErrorObject)	42
− describe/it/test globals (real @types/jest)	30
+ @types/express	21
all fixes	0

Fixed here

• server.ts — the one real bug. provenance.ingestHash is optional by governance (the USS v1.1 schema requires only source/providerId/signalId) but was passed to checkDuplicate/recordIngest, which require string. With AFI_INGEST_DEDUPE=1, the first hashless signal would be recorded under the literal key undefined and every subsequent hashless signal rejected 409 as a duplicate — silently dropping real signals. Guards added; they are no-ops on every payload this route can produce today. This makes D6 ("turn dedupe on?") not a one-line change.
• mergeEnrichedView.ts — three as Record<string, unknown> casts replaced by a declared ReactorEnrichedView interface. Compiles to byte-identical JS.
• cpjMapper.ts — comparator parameters annotated, semantics untouched (feeds generateIngestHash).
• tsconfig.build.json (new) — keeps the compiled test suite out of dist/. CI gains a typecheck step.

Deliberately NOT fixed — needs a decision

ccxt declares OHLCV = [Num × 6] with Num = number | undefined (a venue may return a partial candle) while OHLCVCandle requires number. The code defends ticker.last/timestamp but not ticker.symbol and no candle field — an inconsistently applied assumption on the live Blofin feed. Rejecting or defaulting a partial candle changes scoring behaviour, so it is filed as D8-R2, not fixed.

Zero behaviour change — proven by diffing emitted JavaScript

Tests passing only proves the tests didn't notice. Both branches were compiled and the emitted .js compared directly against a main worktree:

Of 87 emitted files, exactly ONE differs semantically — dist/src/server.js, by exactly the two intentional guard statements. mergeEnrichedView.js and cpjMapper.js are byte-identical, confirming those changes were purely type-level.

• ✅ 670/670 tests, 47/47 suites (baseline is 670, not the 664 in the worklist)
• ✅ provenance golden 312da118… and all 140 fixtures byte-identical
• ✅ compiled honest-unavailable smoke passes (exercises the modified CPJ route)
• ✅ npm ci --dry-run clean; lock diff is @types/express + transitive @types only

Was 324 a floor?

Yes, but only modestly: +10 (→ 334). Three findings were already in the inventory, so the compiler mostly corroborated the manual audit. The real yield was a different class: 2 latent defects on live paths, 9 shims shadowing real third-party types (incl. ajv — schema validation was unchecked), and 9 of 56 test files that never execute (testMatch is an allowlist; 3 have no marker at all, including a 630-line integration suite for the live CPJ route).

Review

4 adversarial lenses + refutation agents. The review changed the diff in six ways — including switching bundler → NodeNext (only NodeNext enforces .js specifiers at compile time; bundler delegated that to a guard CI doesn't run), catching a factual error in the report (afi-core emits 19 files, not 22 — 3 were stale residue), and reverting unrelated lockfile drift. Two "REAL_BUG" calls on mergeEnrichedView were refuted and corrected to NOISE.

Full report: reports/afi-d8-type-boundary-v0.1.md · Register: inventory §3.2

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.

[kilo-code-bot]

Kilo Code Review could not run — your account is out of credits.

Add credits or switch to a free model to enable reviews on this change.

[Gio2050]

⚠️ The red CI here is expected — it is the merge-order coupling, and it confirms the new step works

Validate AFI Reactor fails at the new Typecheck (src + tests, strict) step with exactly one error and nothing else:

src/pipeline/nodes/laneView.ts(77,7): error TS2322: Type 'number | undefined' is not assignable to type 'number | null'.
  Type 'undefined' is not assignable to type 'number | null'.

Cause: this workflow checks out afi-core with no ref: (validate-all.yml:22-26), so CI always builds afi-core's default branch. AFI-Protocol/afi-core#28 — which widens FroggyEnrichedView.technical.indicators to match what this projection actually produces — is not on afi-core main yet.

Fix: merge AFI-Protocol/afi-core#28 first, then re-run this job. Locally, against afi-core with #28 applied, npm run typecheck exits 0.

This is arguably the best possible demonstration that D8 did what it claimed. Before this PR, the reactor was structurally immune to anything afi-core did to its types — the ambient declare module shims meant a cross-repo type change could not fail a build. That immunity is now gone, deliberately, and the very first cross-repo skew it could catch is the one it caught.

Follow-up worth doing separately

afi-core is the only cross-repo build input that records no revision — an unpinned file:../ edge, while afi-infra is pinned by SHA (git+https://…#<sha>). Now that it is a compile-time dependency, recommend either pinning ref: on the sibling checkouts in validate-all.yml, or moving afi-core to the same pinned-git form. Tracked in reports/afi-d8-type-boundary-v0.1.md §7.5.

(Kilo Code Review fails app-side on every AFI PR and is not a signal — judge by repo CI. Scan for Secrets passes on both runs.)