fix(test): close-out audit findings — caching guard never covered EventHandlerEntry

[keithdv]

Follow-up to #73, from the TRIM arc's Step 7 close-out audit. Verdict was CONCERNS with four veto-tier findings; this addresses all four. Full record: docs/todos/TRIM-dto-trimming-preservation-gaps/reviews/close-out-audit.md.

Two of the four were defects in TRIM-006 itself, and both had been self-reported as verified in #73. That is the part worth reading.

V1 — the guard never covered EventHandlerEntry

The fixture declared two matching static handlers for the same event type on one [FactoryEventHandler<T>] class. That is the NF0502 ambiguous-match shape: the transform reports the diagnostic and continues without adding an entry.

So RelayHandlerModel.Entries was empty, and EventHandlerEntry.Parameters / ServiceParameters / AllParameters — the exact three fields #73 converted to EquatableArray<T> — were never constructed and never guarded. The RelayHandler red→green transition in #73 was real, but driven entirely by Usings and Diagnostics; the nested type was never in play. FactoryGenerator.cs:101 returns early on Entries.Count == 0, so the renderer never ran either. The fixture's own comment asserted the opposite.

Fixed with two distinct event types, one matching handler each. Coverage re-proven rather than assumed: reverting EventHandlerEntry.Parameters to IReadOnlyList<T> now turns the RelayHandler case red, which it provably did not before.

Two fixture-health tests added, because "the fixture silently stopped exercising the thing" was the actual failure mode and nothing detected it:

• Fixture_ProducesNoDiagnostics — a diagnostic means some transform took an early-out and never built the model under guard.
• Fixture_EmitsRelayHandlerOutput — the relay output stage returns early on empty Entries, so an empty fixture emits no file.

The first caught a second latent defect on its first run: [Fetch] on an interface-factory member (NF0106 — the interface is the boundary), which had been degrading the FactoryInterface branch too.

V2 — the byte-identity evidence was vacuous

#73 claimed generated output was byte-identical, "verified empirically rather than by inference," on the basis of zero git status drift under Generated/.

.gitignore:405 is **/Generated/ and git ls-files | grep Generated/ returns zero tracked files. Git could not have reported drift regardless of what the generator emitted. The check proved nothing.

Replaced with a real measurement: IntegrationTests' full generated tree emitted with the pre-fix generator (710498c^) into a clean directory, then with HEAD's, diffed recursively — 256 files byte-identical, including all 16 .FactoryEventHandler.g.cs. The conclusion held; the evidence had not.

V3 / V4 — container integrity

• TRIM-007's plan header read In Progress against a Done Index row. Reconciled.
• Plan-review findings B8 (nothing pins the guard's runtime throw) and B10 (16 emission assertions can pass vacuously via naive IndexOf slicing) had no disposition anywhere. Both now carried in a Deferred Work Carrying Forward table with explicit accept-or-queue reasoning.

Also: DtoTypeWalker's header named one caller when it now has three (this file's stale header was itself a TRIM-001 finding), and the TRIM-007 log entry's imprecise "Amend" decision is corrected in place per the append-only rule.

Tests

603+603 unit (up 2 — the new health tests), 561+561 integration, 86+86 Design. 0 failures across net9.0 and net10.0.

Note on the release

The v1.7.0 release is deliberately held — not forgotten. docs/trimming.md:35 claims [Execute] static factories' registration lambdas are trimmed away while :222 accurately describes the registrar-DAM mechanism that makes that false. Shipping today would publish IP-protection guidance the arc itself disproved. The release reopens once the registrar-DAM fix lands. Cost is recorded plainly in the todo: zTreatment PCB-003 stays blocked meanwhile.

🤖 Generated with Claude Code