rdmsm4x-changelog-20260914-1702-tyrell-event-key-diagnostics

Completed isolated Tyrell event-key diagnostics and relayed a correctness proposal to Claude and the canary coordinator.

Tyrell complementary diagnostics — 2026-09-14

Rich explicitly directed coordination with Claude and continued iteration. Claude's observed message identifies an app-only UI lane (Build22); this packet is a separate copied-source TyrellCore diagnostic proposal. The coordinator owns Build23 and the protected rdmbair15m5 canary. No production checkout, installed bundle, launch agent, database, or canary was changed by this diagnostic work.

Proven correctness finding

mergeFleetEvents concatenates timestamp, host, kind and detail with colons to deduplicate events. Distinct string-field boundaries can collide: kind custom:scan plus detail file equals kind custom plus detail scan:file. The actual copied production actor retains only one of these two events. The proposed structured Hashable key retains both and still removes an identical duplicate. It leaves the persistence queue empty for relayed events.

Source provenance: commit 093d653493be92aed0fb39779f52d5eaef04230b; source-provenance.json records the extraction. event-key-collision.patch changes only the actor key representation. Timestamp remains String(event.ts), preserving the former textual timestamp identity. This is a review proposal, not integrated or release-tested.

Reproduce from an authorized Aqua Terminal: zsh /Users/richh/dev/_handoff/tyrell-allocation-attribution-20260914/run_collision_check_v1.zsh Recorded collision-20260914-165758: source.exit=1 (one distinct event), variant.exit=0 (two distinct events), runner.exit=0 (expected differential observed).

Memory evidence does not justify a cure claim

Eight sequential V2 runs in source/variant/variant/source order, each steady and burst, completed with comparison runner.exit=0. Each runs 120 JSON decode/merge/encode cycles with 5000 synthetic events. No real activity data is used. Idle physical footprint before the optional diagnostic relief call:

The variant is lower for steady input but higher for burst input. Live malloc bytes are approximately 12.9 MiB for both. Fixture generation, JSON, allocator history, and process runtime are all present; this short synthetic test cannot attribute the live canary's trajectory. All eight pressure-relief probes reported zero bytes released. No production pressure-relief call is proposed.

Reproduce: zsh /Users/richh/dev/_handoff/tyrell-allocation-attribution-20260914/compare_activity_replay_v1.zsh Exact copied inputs, JSONL measurements, stderr and runner log: compare-20260914-165729/. The earlier run-20260914-165545 used V1 with an extra retained fixture; do not compare its absolute footprint against V2 as a treatment effect.

Disposition

Recommend review of the collision correction as a separate correctness change after the frozen Build23 canary. Do not alter the current candidate, gate, warmup or protected host for this finding. Full repository tests and integration review are still required before shipping this proposal. Build23 addresses a separate proven pending-event queue leak; its memory acceptance remains the coordinator's decision.

All files are isolated under this packet directory; undo is to leave the proposal unapplied. No shared library change is warranted by this actor-local finding. No additional subagents were spawned for this diagnostic pass.

Coordination check-in updated: /Users/richh/.agent-coordination/checkins/codex-tyrell-allocation-attribution-20260914.json Bus message: 20260914-170210-AE5EDFA7. Existing Build23 packet and canary left untouched. Local reproduction scripts compiled copied Swift source using xcrun swiftc in Aqua Terminal; eight comparison runs and differential collision test completed. No installed state changed; proposal may be left unapplied.