TCF audio lifetime: local measurement note
Measured: 2026-09-28

Extracted summary of private test artifacts. This is not a standalone reproducible full benchmark or a public code release. No new measurements were made for this export.

REVISION
base_head: abe15eb9918fa08648978d9822fd9dfcf6004e55
state_at_measurement: local uncommitted/unmerged/undeployed candidate
changed_runtime: backend/application/eval_t3_followup.py
baseline_runtime_sha256: e08ad590f575e25737da6a383e840c3df1f4d88702256d509673464fa7a2a169
candidate_runtime_sha256: e52ac13cbf450fa07c3224a5c96f008deec1a30d8e42254405c5dd06385a3f17

ENVIRONMENT
runtime: Python 3.12.14 on Linux
database: Real, isolated local PostgreSQL 16.14; application Alembic revision 20260831_0021
cpu_quota: 2
memory_limit_bytes: 3221225472
host: Shared WSL2 host; no reserved CPU
concurrency: One worker job at a time per fixture
warmup: No separate warm-up

METHOD
fixture_bytes: 4194304
fixture: 4 MiB of synthetic bytes, not playable audio and not a real user's recording.
shared_true: The DSP stub returns the original buffer.
shared_false: The DSP stub returns a distinct 4 MiB cleaned buffer.
real_boundary: Durable encrypted job, local database claim, actual worker decode/hash verification, dispatch, application use case, and conversation/job persistence.
substituted_boundaries: Audio validation/DSP, speech-to-text, and examiner are stubs. No real speech processing or hosted AI inference is measured.
sampling: For each state, three fresh pytest processes ran the same ordered 12 success cases. Each reported case has three samples across those processes. A separate fresh process was not started for each individual case.
reported_cases: Eight examiner-boundary cases per state: T1/T3, start/non-final follow-up, shared/distinct buffers. Four final-turn success cases did not invoke the examiner and have no measurement rows.
heap: Current Python allocations traced by tracemalloc at the first examiner call, after an event-loop yield and gc.collect(). Tracing starts after the job claim and before the actual worker call; fixture setup is outside this interval.
peak: Peak Python allocations traced from that tracing start up to the same examiner observation. It is not a whole-process or full-service peak.
statistics: Exact bytes: median, minimum, and maximum of three observations per case and state. 1 MiB = 1,048,576 bytes.
excluded_process_metrics: RSS/PSS values are intentionally excluded. Baseline assertion failures may retain traceback state into later cases in the same pytest process, contaminating the before/after comparison. These runs do not establish a causal RSS/PSS reduction.

VERIFICATION
measurement_baseline: Each of three baseline runs: 8 retained-heap assertion failures, 4 final-turn cases passed. Conversation/job assertions completed before the heap failures.
measurement_candidate: Each of three candidate runs: all 12 success cases passed.
oracle: Relevant heap and content assertions were unchanged after the failing baseline. Import/fixture lint cleanup and omission of untraced retry observations did not weaken those assertions.
related_regression: 177 passed, 1 warning in 60.21s; no skipped tests. This is the related suite, not the full repository suite.
regression_utc: 2026-09-28T11:34:08.614171+00:00
cancelled_start_limit: Cancelled-start cases verify preservation of durable processing state and retry data. They do not establish full automatic restart recovery through the real reconciler and database. Reconciliation unit tests substitute the database.
not_established: Real-provider E2E, speech/model quality, production capacity, steady-state service RSS, and invoice savings are unverified. The candidate was local, uncommitted, unmerged, and undeployed at measurement time.
export_integrity: Exported metric dictionary exactly equals the manifest's complete heap/peak selection and independently recomputed statistics from the six stored measurement receipts.

EXACT MEASUREMENTS (bytes)
case | heap median | heap min | heap max | peak median | peak min | peak max
baseline/t1/start/shared=False | 4591861 | 4591346 | 4595436 | 23960011 | 23943791 | 23984680
baseline/t3/start/shared=False | 4272973 | 4272403 | 4273451 | 23843060 | 23826156 | 23843592
baseline/t1/start/shared=True | 4271155 | 4271105 | 4271209 | 23825532 | 23824080 | 23833728
baseline/t3/start/shared=True | 4270885 | 4270885 | 4271409 | 23842248 | 23833528 | 23849916
baseline/t1/followup/shared=False | 4296599 | 4296543 | 4296655 | 23833630 | 23833630 | 23841822
baseline/t3/followup/shared=False | 4247019 | 4246969 | 4248603 | 23826430 | 23825374 | 23841758
baseline/t1/followup/shared=True | 4247115 | 4247111 | 4247115 | 23833678 | 23833674 | 23833678
baseline/t3/followup/shared=True | 4246943 | 4246935 | 4246943 | 23833474 | 23825282 | 23841658
candidate/t1/start/shared=False | 400892 | 398296 | 401542 | 23944064 | 23943489 | 23953246
candidate/t3/start/shared=False | 78078 | 78074 | 79658 | 23834344 | 23827736 | 23858920
candidate/t1/start/shared=True | 77370 | 77366 | 77412 | 23834256 | 23826060 | 23842452
candidate/t3/start/shared=True | 76660 | 76614 | 77188 | 23841720 | 23833528 | 23850440
candidate/t1/followup/shared=False | 103898 | 103202 | 104300 | 23843402 | 23826490 | 23843406
candidate/t3/followup/shared=False | 53210 | 53210 | 53738 | 23834094 | 23826430 | 23834094
candidate/t1/followup/shared=True | 53830 | 53306 | 54586 | 23842398 | 23834206 | 23842922
candidate/t3/followup/shared=True | 53654 | 53608 | 53886 | 23833474 | 23826330 | 23842718

SOURCE FINGERPRINTS
manifest.json SHA-256: e23255abe8827323cb362f69c3718caaadebd8f247fdd76454cea3f8b0003372
baseline1.json at 2026-09-28T11:25:43.468301+00:00 SHA-256: 343688ca22f80a7a9ddc68bfc8e7996916ba841402972dbf20ca3b36ad9ac16a
baseline2.json at 2026-09-28T11:26:16.102798+00:00 SHA-256: 36807a8d26d50426a454472e846b93a5b93859ada79935134b305ae76cf33796
baseline3.json at 2026-09-28T11:26:42.083695+00:00 SHA-256: 4470c35b89aa009d2bc78926f58e4d0dbf664429e0651d4cb5371750da2c2586
candidate1.json at 2026-09-28T11:27:25.593254+00:00 SHA-256: b17882cd8c2fbab8804e88e168795783ea4716fed9ddc38e5f5d2ae55113560d
candidate2.json at 2026-09-28T11:28:33.465530+00:00 SHA-256: e55766fa7d70570ce8cdbc5265d947fb0dcdd6b4aab8036a9dd0673c0c4fcd80
candidate3.json at 2026-09-28T11:29:05.348266+00:00 SHA-256: 2c5f24f0f6aa51a5905e8366a9d3963ff13522e3ab859ed67ca65699369a2965
regression-final.json SHA-256: 41ca15c371bb5b35b8c43a2e0cdba2a1de01df7679db2610f228c14037f0cf7c