docs(01): create verified fidelity baseline plans

This commit is contained in:
UnixNight committed 2026-11-12 17:38:34 -05:00
1 parent 36f24e4474
commit cf64b8ce18
21 files changed
+1473 -32

No files matched your search

@@ -0,0 +1,105 @@
---
phase: 01-reproducible-fidelity-baseline
plan: "06"
type: execute
wave: 4
depends_on: ["01-05"]
files_modified:
- tools/recon/fidelity_capture.py
- scripts/collect-fidelity-linux.sh
- tests/test_fidelity_capture.py
- docs/evidence/linux-session.md
autonomous: true
requirements: [EVID-01, EVID-04]
estimate:
tokens: 25000
raw_tokens: 25000
tasks: 2
confidence: low
must_haves:
truths:
- "An identified native Linux package yields actual input/event and audiovisual captures where execution is available; synthetic/headless replays remain distinct (EVID-04)."
- "Continuous route and focused captures cover all natural/skipped, fresh/retained and ordinary/edge cells per D-01–D-04, with missing executions explicitly open."
- "Physical input/audio and real visible gameplay require actual native receipts/human review; dummy audio and window-only fixtures cannot complete acceptance."
artifacts:
- path: tools/recon/fidelity_capture.py
provides: "Bounded native Linux capture orchestration and result collection"
- path: docs/evidence/linux-session.md
provides: "Native Linux route protocol, capture setup and physical review agenda"
- path: scripts/collect-fidelity-linux.sh
provides: "Local identified-package launch/collection entry point"
key_links:
- from: tools/recon/fidelity_capture.py
to: "Identified Linux checkpoint package and port observer"
via: "Actual process invocation, accepted input/events, raw media and receipt identity"
- from: "Native Linux receipts"
to: "Paired review and final acceptance audit"
via: "Validated case/session/package/media identity and explicit missing cells"
---
<objective>
Collect current native Linux baseline evidence using the completed checkpoint rather than substituting synthetic fixtures for ordinary gameplay. Prepare all physical/audible review items for the same final combined review.
Artifacts this phase produces: fidelity_capture.py preflight/run/collect CLI; collect-fidelity-linux.sh; tests/test_fidelity_capture.py; linux-session.md; ignored artifacts/fidelity/linux sessions containing receipts, logs, raw video/audio, observer events, annotations, media probes and images. Original-side capture remains on Windows through DxWnd.
</objective>
<execution_context>
@/home/cristian/.codex/gsd-core/workflows/execute-plan.md
@/home/cristian/.codex/gsd-core/templates/summary.md
</execution_context>
<context>
@AGENTS.md
@.planning/ROADMAP.md
@.planning/phases/01-reproducible-fidelity-baseline/01-CONTEXT.md
@.planning/phases/01-reproducible-fidelity-baseline/01-01-SUMMARY.md
@.planning/phases/01-reproducible-fidelity-baseline/01-04-SUMMARY.md
@.planning/phases/01-reproducible-fidelity-baseline/01-05-SUMMARY.md
</context>
<tasks>
<task type="tracer" tdd="true" id="01-06-01">
<name>Capture one actual Linux focused moment and import its observed evidence</name>
<files>tools/recon/fidelity_capture.py, scripts/collect-fidelity-linux.sh, tests/test_fidelity_capture.py, docs/evidence/linux-session.md</files>
<read_first>docs/evidence/scenarios.json; tools/recon/fidelity_receipts.py; scripts/test-window-transitions.sh; docs/evidence/windows-session.md; 01-RESEARCH.md capture/media guidance; actual completed Linux package BUILD-INFO and manifest.</read_first>
<behavior>At least 10 tests exercise capture launch→probe→hash→import with bounded process/media doubles, real receipt validation and explicit fixture category; missing display/audio, recorder exit, timeout/abort, stale package, empty/frozen media indicators and native-versus-headless classification remain incomplete.</behavior>
<action>Create tests first. Implement bounded orchestration that verifies completed package/resource identities, records OS/GPU/driver/Vulkan/SDL and capture-tool versions, launches the native Linux package using normal controls plus passive observer and collects external recording and actual logs. Use available local FFmpeg/OBS capabilities only after preflight; record the selected X11/Wayland/display/audio capture methods rather than assuming Pulse/PipeWire availability. Preserve raw video/audio and any timestamp-preserving remux preview; ffprobe validates actual streams/durations, while human visibility/audibility remains separate. Prove one focused normal-play dialogue-return/neutral-input moment end to end through launch, logs/events, media collection, import and viewer links. Existing SDL injection may support a clearly labeled automated native replay, but cannot be relabeled physical user input. A smoke or headless fallback stays fixture evidence. If this environment cannot run/capture native audio/video, publish its exact failed preflight/open receipt and finish the usable collector/procedure without blocking remaining bundle preparation. This conditional is an execution dependency, not a waiver of native acceptance. Do not ask the user for an early session.</action>
<acceptance_criteria>Counted capture round-trip tests pass; either an actual identified native focused receipt or a precise failed-preflight receipt exists; no synthetic/automated/dummy-device evidence claims physical/audible acceptance.</acceptance_criteria>
<verify>
<automated>python3 -c 'import sys,unittest; s=unittest.defaultTestLoader.discover("tests",pattern="test_fidelity_capture.py"); assert s.countTestCases() >= 10,"fewer than 10 capture tests"; sys.exit(not unittest.TextTestRunner(verbosity=2).run(s).wasSuccessful())'</automated>
<fails_when>Nonzero exit, fewer than 10 tests, absent round-trip output or an unavailable/synthetic capture promoted to native acceptance.</fails_when>
<human-check>The final combined review checks actual Linux physical input, moving game content and audible output, including the observer's recorded scope and any capture overhead.</human-check>
</verify>
<done>One focused Linux capture path is usable, and its actual execution result or precise access limitation is represented honestly.</done>
</task>
<task type="auto" tdd="true" id="01-06-02">
<name>Collect the complete Linux journey matrix with preserved entry lifetimes</name>
<files>tools/recon/fidelity_capture.py, scripts/collect-fidelity-linux.sh, tests/test_fidelity_capture.py, docs/evidence/linux-session.md</files>
<read_first>Task 01-06-01 collector; docs/evidence/scenarios.json; docs/gameplay/main-menu-player-lifetime.md; docs/gameplay/jimmy-movement-stabilization.md; current checkpoint package identities.</read_first>
<behavior>Extend to at least 14 tests proving full case enumeration, route continuity, focused obligations, fresh/retained separation, natural/skip identification, interrupted matrix resume without reusing sessions and explicit unavailable cells.</behavior>
<action>Expand the real collector/procedure to all required continuous/focused matrix cases per D-01–D-04. Each continuous lab→VR04→lab→outside recording preserves actual scene/control/save/RETURN lifetimes; independent focused setups are labeled and cannot replace the continuous journey. Keep ordinary play separate from held-through-dialogue/return, banked jumping left/right, landing and idle edge runs. Capture raw elapsed/event deltas and annotation uncertainty without inventing tolerances or repairing bank/movement. Use normal gameplay; debug warps/grants and smoke route fixtures remain narrow supplementary evidence. Verify source originals before/after the capture checkpoint. Attempt all runnable native Linux cells, record actual failures and missing media/physical/audible checks, and append remaining local Linux review actions to the single final session agenda. Re-run only failed/incomplete cells under new session IDs with retained history. Export receipt/review data for the final acceptance audit; do not update capability/root status files owned by final integration.</action>
<acceptance_criteria>All eight cells and their continuous/focused obligations are represented by actual receipt IDs or precise open dependencies; no count-only completeness or disconnected scene substitution passes.</acceptance_criteria>
<verify>
<automated>python3 -c 'import sys,unittest; s=unittest.defaultTestLoader.discover("tests",pattern="test_fidelity_capture.py"); assert s.countTestCases() >= 14,"fewer than 14 capture tests"; sys.exit(not unittest.TextTestRunner(verbosity=2).run(s).wasSuccessful())'</automated>
<fails_when>Nonzero exit, fewer than 14 tests, or a missing cell/lifetime/continuity assertion fails.</fails_when>
<automated>python3 tools/recon/inventory.py ../Original --verify docs/original-manifest.json</automated>
<fails_when>Nonzero exit or missing Verified 539 unchanged original files.</fails_when>
</verify>
<done>The Linux evidence matrix is collected as far as actual access permits, with all remaining acceptance explicitly retained for final review.</done>
</task>
</tasks>
<threat_model>
ASVS level 1; high/critical unresolved threats block acceptance.
| Boundary | Description |
|---|---|
| Capture configuration → owned processes | Bound selected recorder/native executable and output paths |
| Captured artifacts → receipt | Real media/tool/input provenance must survive import |
| Threat ID | Category | Component | Severity | Disposition | Mitigation Plan |
|---|---|---|---|---|---|
| T-01-16 | Tampering | Native collector | high | mitigate | Argument arrays, contained fresh output and manifest checks before launch/import |
| T-01-17 | Spoofing | Linux evidence classification | high | mitigate | Actual native/input/audio method fields, fail-open acceptance prohibited, explicit unavailable cells |
| T-01-18 | Denial of service | Recording process | medium | mitigate | Bounded durations, owned-process shutdown and incomplete receipts on abort |
| T-01-SC | Tampering | Package installation | high | mitigate | Existing local capture tools only; no new packages |
</threat_model>
<verification>Counted collector tests and original preservation must pass. Native evidence gaps stay open regardless of tooling success. Actual physical input/audibility is reviewed once with the final combined materials.</verification>
<success_criteria>Current Linux route evidence is reproducible and precisely classified, with no unsupported native acceptance claim.</success_criteria>
<output>Create .planning/phases/01-reproducible-fidelity-baseline/01-06-SUMMARY.md with actual case receipt IDs, missing cells and capture limitations.</output>