The gate tripped on a uniformly slow runner: every benchmark ran at
~2/3 of the previous CI run and nostrInbound.duplicate fell to 87% of
its floor. Root cause: floors were derived from local numbers, but CI
slowdown is benchmark-dependent - sub-millisecond passes amplify runner
overhead (the duplicate path runs at ~20% of local speed on CI while
most benchmarks run at 40-60%). Floors are now ~50% of the slowest
observed CI run, recorded alongside the local references. Every floor
remains 10-200x above known regression values (the pre-optimization
duplicate path measured 2.2k/sec against the 250k floor), so order-of-
magnitude regressions still fail loudly. Verified against the slow
run's numbers: all 11 pass.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
ConversationStore.auditInvariants() verifies the per-conversation and
store-level message-ID indexes, caps, timestamp ordering, unread-set
membership, and selection validity - wired to the existing read-receipt
cleanup cadence, loud (.error) on violation, sampled heartbeat when
healthy (~2.8ms per audit at 5k messages, benchmarked and floored).
Router drops log both outcomes (marked failed / skipped by no-downgrade
guard); relay cap evictions, age sweeps, and jittered reconnect delays
log their counts; mirrored republishes get a sampled proof line. 11 new
invariant tests corrupt store state through DEBUG-only hooks since the
single-writer lockdown makes those states unreachable via intents.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Package.swift's .process("Noise") resource claim silently excluded all
of bitchatTests/Noise/ from compilation since Oct 2025 - including a
complete official-vector runner (cacophony + snow XX transcripts,
transport messages, handshake hash, byte-identical to upstream).
Narrowing the resource to the JSON file and loading via Bundle.module
brings 51 Noise tests back to life, with a guard asserting each
vector's protocol name matches the app's.
CI gains a performance floor gate: perf-floors.json carries deliberately
generous floors (~25% of measured throughput) that catch algorithmic
regressions without flaking on runner variance; PERF lines reach the
gate via an O_APPEND side-channel file since swift test --parallel
swallows passing tests' stdout.
Tests are now hermetic: FavoritesPersistenceService uses an in-memory
keychain under test (fixes the securityd hang that blocked pipeline
benchmarks locally) and read-receipt persistence uses a wiped scratch
UserDefaults suite instead of .standard.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
jack