Files
Neutron/.planning/phases/01-reproducible-fidelity-baseline/01-06-SUMMARY.md
T

201 lines
11 KiB
Markdown

---
phase: 01-reproducible-fidelity-baseline
plan: "06"
subsystem: testing
tags: [linux, fidelity, capture, ffmpeg, provenance]
requires:
- phase: 01-01
provides: Typed receipts and eight-cell scenario contract
- phase: 01-02
provides: Local paired review export
- phase: 01-04
provides: Passive runtime observer
- phase: 01-05
provides: Identified sealed Linux checkpoint package
provides:
- Bounded local Linux preflight/run/collect and immutable receipt import
- Forty Linux obligations with exact open dependencies and verified attempt history
- Actual host failed-preflight receipt and local review export
affects: [01-07]
tech-stack:
added: []
patterns: [fresh output allocation, owned child deadlines, sealed package checks, explicit open acceptance]
key-files:
created:
- tools/recon/fidelity_capture.py
- scripts/collect-fidelity-linux.sh
- tests/test_fidelity_capture.py
- docs/evidence/linux-session.md
modified: []
key-decisions:
- "A real host with no display and only auto_null audio yields failed native preflight, never synthetic replacement gameplay."
- "One common failed environment dependency can explain all forty open Linux obligations without inventing forty game launches."
- "Retry history revalidates actual saved receipts and requires new launch/session IDs; focused captures never substitute for the continuous journey."
requirements-completed: [EVID-01, EVID-04]
requirements-acceptance: "OPEN: conditional tooling/preparation contributions complete; actual native physical/audible acceptance remains 0/40 Linux and 0/120 overall. Shared IDs remain Pending while plan 01-07 is unfinished."
coverage:
- id: linux-collector
description: Bounded process/media doubles, actual receipt import and classification
requirement: EVID-04
verification:
- kind: integration
ref: tests/test_fidelity_capture.py
status: pass
human_judgment: false
- id: linux-matrix
description: Eight cells and forty continuous/focused obligations with retained/fresh history
requirement: EVID-01
verification:
- kind: integration
ref: artifacts/fidelity/linux/01-06-matrix.json
status: pass
human_judgment: false
- id: actual-native-capture
description: Actual moving gameplay, physical input, real audible output and original/native comparison
verification: []
human_judgment: true
rationale: "Actual host preflight found no display session and only auto_null audio. All physical/audible and full-route obligations remain assigned to the final combined session."
actuals:
tokens: 14371
tasks: 2
commits: 4
basis: "ceil(chars/4) over the realized four-file production/test/procedure diff: 57484 characters; persisted ledger measures four task commits before summary metadata."
plan_head_before: 31ae6a6ab2ab1c342361de376a8756dafafb8227
plan_head_after: 0172916ffd9e3120e57c87c3bd853965f3e4d196
duration: 14 min
started: 2026-11-13T16:49:00Z
completed: 2026-11-13
status: complete
---
# Phase 01 Plan 06: Linux Capture and Open Native Matrix Summary
**A bounded FFmpeg collector preserves package identity, failed native preflight and all forty Linux obligations without substituting synthetic gameplay.**
## Accomplishments
Both conditional plan tasks are complete. The collector provides `preflight`,
`run`, `collect` and `matrix`; the shell launcher does no build, download or
dummy-device fallback. It validates the sealed existing package/resources,
records OS/tools/hash/device/GPU information, and supports installed X11/Pulse
recording with normal `NewGame.tsk --walk` entry and passive observer. Actual
runtime SDL/active GPU/Vulkan versions remain unavailable without a game launch.
Fresh contained output, inert argument arrays, selected local sources, bounded
metadata/log/media sizes, one deadline and owned-child cleanup implement the
plan's process/filesystem trust boundaries. Raw lossless FFV1/PCM MKV, media
probes, logs, clocks and prelaunch identities survive collection/import. Media
emptiness, missing streams, frozen indicators, interrupted/failed children and
post-launch package changes keep observations incomplete. No automatic acceptance,
retiming, volume adjustment or gameplay/movement change was introduced.
The matrix enumerates each lifetime/dialogue/input cell and all five obligations,
including exact events/procedures and continuous lab → VR04 → lab → outside.
Saved retry history is hash-checked; a new attempt needs a new launch/session ID.
Annotations bind contained prerequisite recordings and actual copy-state identity
records without promoting them to physical/audible review. The Linux procedure
joins the single later Windows/original combined agenda.
## Task Commits
1. Task 1 RED: `9c59f60` — require bounded Linux capture round trip.
2. Task 1 GREEN: `16f5e12` — collect bounded Linux capture and honest failed preflight.
3. Task 2 RED: `e17cfbe` — require full Linux matrix continuity and resume history.
4. Task 2 GREEN: `0172916` — preserve full Linux obligation matrix and capture history.
Normal hooks and the supplied-root/protected-branch checks ran. The primary
explicitly authorized recovery after my first guard extraction omitted its bound
assignment; the complete literal guard then passed before any write. No root drift,
tracked deletion, runtime edit or unrelated `tools.yaml` change occurred. Primary
owns integration reports and push.
## Actual Execution and Package Identity
The outside-sandbox host probe found a TTY session, no `DISPLAY`/`WAYLAND_DISPLAY`
or X11 socket, no resolved Xvfb/OBS command and only suspended PipeWire Pulse
`auto_null`/`auto_null.monitor`. FFmpeg/ffprobe are installed distribution version
`6.1.1-3ubuntu5+esm15`; local help confirms the selected frozen-frame filter.
`lspci` succeeded; `vulkaninfo` is absent. The exact observations/tool hashes and
selected settings are retained. No game/recorder was launched, and no raw
video/audio or observed gameplay/input event was fabricated.
The existing package `artifacts/fidelity/01-05-packages/linux-x64` was reused,
without publish or bulk copying. Its file-list hash remains
`124cf271e4952f3822eec5c24358447fcffc9e9e1803aded8d96d4d4472dc5a3`.
Its actual source is `446930d2a30d6d616c60ac5e59e4169a840bea09`, dirty yes,
diff `90cef95669caece53fe2d2c4736cf2ac05ed27f5d25ca03ab572589686bcb73d`;
the collector revision never replaces that runtime provenance.
| Evidence | Exact local result |
|---|---|
| Task-1 diagnostic | `artifacts/fidelity/linux/01-06-focused-preflight/receipt.json`, ID `linux-01-06-focused-preflight-linux`, failed/open |
| Current typed dependency | `artifacts/fidelity/linux/01-06-matrix-preflight/receipt.json`, ID `linux-01-06-matrix-preflight-linux`, failed/open |
| Current applicable import | `artifacts/fidelity/linux-current-receipts/linux-01-06-matrix-preflight-linux/receipt.json` |
| Full Linux checklist | `artifacts/fidelity/linux/01-06-matrix.json`, 8 cells / 40 obligations, all open |
| Current shared audit | `artifacts/fidelity/linux/01-06-audit.json`, accepted 0 / required 120, invalid `[]` |
| Local review | `artifacts/fidelity/linux/01-06-review/index.html`, failed receipt with no fabricated media |
The older task-1 import remains diagnostic history; select the current typed
receipt for scenario audit. The focused failure is an actual preflight result,
not an executed gameplay moment. Every continuous/focused slot retains the exact
current dependency receipt plus its own actual capture/lifetime/human-review
requirement. No native matrix cell was runnable after the shared preflight failed.
## Verification
- Final counted capture verification: **23 tests pass**, exceeding both 10/14
thresholds. Doubles stay synthetic, native automation stays `port_fixture`,
positive retained file/clock binding and invalid retained-session cases pass.
- Broader `test_fidelity*.py` regression: **156 tests pass** in 36.316 seconds,
including local viewer sockets and existing real .NET/observer fixtures outside
sandbox restrictions. This ran before the final capture-size bound regression;
the final focused suite passed all 23 after that bounded collector change.
Complete report: `artifacts/fidelity/linux/01-06-fidelity-tests.log`.
- Shell syntax and `git diff --check` pass. Both actual failed receipts validate
through the shared validator/importer. The current scenario audit has no invalid
receipts and grants no acceptance.
- Original inventory verified **539 unchanged files** before capture preparation
and repeatedly after collection/final changes. No generated package, original,
trace or raw media is staged.
- Task-1 tracer verification reran and passed before matrix expansion. No blanket
fidelity tolerance or native Windows/original execution claim is made.
## TDD Gate Compliance
Actual records `artifacts/fidelity/tdd/01-06-01-red.json` and
`01-06-02-red.json` both return `RED_EVIDENCE_OK`. Task 1's named target executes
the planned missing capture-API assertion, without import/setup failure. Task 2's
named target executes the missing forty-obligation matrix assertion while the
first twelve tests remain green. Both RED commits precede their GREEN commits.
No separate refactor was needed.
## Deviations from Plan
None in scope. The plan's explicit unavailable-native conditional was exercised:
precise failed preflight, usable collector/procedure and complete open matrix
replace unavailable execution, without waiving acceptance. Additional correctness
checks preserve prelaunch identity after package tampering, bind annotation files
and exclude oversized raw media from unbounded import. These implement the
planned T-01-16/17/18/SC mitigations; no new trust surface was introduced.
## Remaining Acceptance and Next Phase
All **40 Linux** and **120 total** actual original/native acceptance obligations
remain OPEN. A usable display, real output/monitor device, physical play, continuous
and focused recordings, audible/visible/browser review and original-DxWnd/native
Windows comparisons are still required in the final combined session. Independent
fresh execution-copy staging and actual same-process retained state remain visible
procedure obligations; no source package is silently relabeled as a fresh copy.
The unavailable physical/native check is recorded as open WINDOWS ledger entry 2
(`unrun-verify`), not as a passed test or waived truth. No implementation stub or
skipped mandatory automated verification remains. Shared EVID-01/EVID-04 IDs stay
Pending because unfinished plan 01-07 owns their remaining gates.
## Self-Check: PASSED
All four implementation/procedure files exist; all four recorded task commits
are ancestors of HEAD. Actual failed/current receipt files, matrix, audit and local
review export exist. The persisted ledger measures four commits over the recorded
before/after range. Original verification, final counted tests and current receipt
audit pass within their stated scope.