.gitignore

# ── Personal health data & large raw recordings ─────────────────────────────
# uploads/ holds REAL biosignal recordings (overnight SpO2/ECG/PPG, CGM glucose,
# HRV exports, CPAP EDFs). Raw recordings are personal medical data AND large
# (hundreds of MB per night) — never publish them.
#
# EXCEPTION (fixes the fresh-clone gate hole): verify-provenance.html GATE B
# audits committed DERIVED node-export JSONs (summary metrics only — NO raw
# signal samples). These MUST be tracked or GATE B has nothing to read on a
# clone. They are re-included BY EXACT NAME below — never a blanket *.json,
# which would sweep in the dozens of OTHER real-recording summary exports in
# uploads/ that are NOT gate fixtures and are not chosen for publication.
# NOTE: device serials (Polar H10/Verity) were scrubbed from ALL tracked files
#   2026-06-29 (→ placeholders); committed fixtures are provenance-null/clean.
#   Recording DATES remain (own data, low-identifiability, needed for fixtures).
#   Future exports should scrub filenames at the ingest boundary — EXPORT-IDENTITY-BRIEF.md.
uploads/*

# PAPER-ODI4-REPRODUCIBILITY §4 — the SubjectA pilot corpus behind papers/odi4-ahi-bias.html.
# SYNTHETIC (synth-gen, seed 424242, VERSION synth-gen/2.1) so there is NO privacy bar, unlike the real
# recordings this file otherwise excludes. Committed as BYTES, not as a regeneration recipe: §4 allowed
# either, but a recorded seed alone is effectively what existed before — and it failed silently when the
# generator moved, which is why Table 1 stopped reproducing and its corpus "cannot be recovered". Bytes
# cannot move without a diff.
!uploads/O2Ring S 2100_20260511231000.csv
!uploads/O2Ring S 2100_20260512235500.csv
!uploads/O2Ring S 2100_20260513225000.csv
!uploads/O2Ring S 2100_20260514230500.csv
!uploads/O2Ring S 2100_20260515232000.csv
!uploads/ground_truth_night1.json
!uploads/ground_truth_night2.json
!uploads/ground_truth_night3.json
!uploads/ground_truth_night4.json
!uploads/ground_truth_night5.json

# verify-provenance GATE B fixtures (== FIXTURE-PROVENANCE.json "fixtures" keys):
!uploads/OxyDex_2026-06-13_1056_summary.json
!uploads/OxyDex_2026-06-25_0439_summary.json
!uploads/PulseDex_2026-06-25_equiv.node-export.json
!uploads/PulseDex_2026-06-25_events.node-export.json
!uploads/HRVDex_2026-06-25_equiv.node-export.json
!uploads/HRVDex_2026-06-25_events.node-export.json
!uploads/GlucoDex_2026-06-27_equiv.node-export.json
!uploads/PpgDex_2026-06-27_equiv.node-export.json
!uploads/ECGDex_2026-06-27_equiv.node-export.json
!uploads/cpapdex-2026-06-12.node-export.json
!uploads/cpapdex-2026-06-16.json
!uploads/cpapdex_synthetic_golden.node-export.json
!uploads/cpapdex_synthetic_multinight_golden.node-export.json

# CPAPDex BINARY-EDF equivalence leg (CPAP-REAL-CORPUS §P2). These five .edf files are
# SYNTHETIC (tools/make-synthetic-edf.mjs) — closed-form waveforms, no recording of any person,
# header identity fields blank. They are the ONLY committed equiv INPUT in the repo, which is
# what lets the compute()-equivalence diff actually RUN IN CI (every other node's input is a real
# recording, so it is gitignored and its leg skips on a fresh clone). Regenerate byte-identically
# with `node tools/make-synthetic-edf.mjs`.
!uploads/20260613_231433_BRP.edf
!uploads/20260613_231433_PLD.edf
!uploads/20260613_231433_SA2.edf
!uploads/20260613_231433_EVE.edf
!uploads/20260613_231433_CSL.edf
!uploads/cpapdex_synthetic_edf_golden.node-export.json
# CPAPDex live-vs-SD comparator pin-twin — SYNTHETIC BRP pair (one wall-clock flow sampled into two
# files, sd −159 s ×0.998) reproducing the night pin's identity verdict + 27.35-min overlap geometry.
# Real patient slices stay gitignored (health data, public repo); only the synthetic twin is committed.
!uploads/cpapdex_comparator_live_twin_BRP.edf
!uploads/cpapdex_comparator_sd_twin_BRP.edf
!uploads/cpapdex_comparator_golden.json

# P9 — SYNTHETIC vendor-format inputs + their goldens (tools/make-synthetic-inputs.mjs).
# The equivalence gate (compute() ≡ committed export) did not run in CI AT ALL: every node's
# input was a real recording, hence gitignored, hence ⊘ skipped on a fresh clone. These carry no
# personal data, so they ship — and the gate finally has teeth in CI. Deterministic: regenerate
# byte-identically with `node tools/make-synthetic-inputs.mjs`.
!uploads/synthetic_oxydex_o2ring.csv
# ADVERSARIAL twins (DEEP-AUDIT-2026-07-11 §1/§8/§9). Every input above is CLEAN, short, DMY and
# gapless, so three of the audit's worst defects were STRUCTURALLY INVISIBLE to CI. These express the
# shapes that caught them: an MM/DD date order, a dropped-row night, a full-length overnight.
!uploads/synthetic_oxydex_o2ring_dmy.csv
!uploads/synthetic_oxydex_o2ring_mdy.csv
!uploads/synthetic_oxydex_o2ring_lossy.csv
!uploads/synthetic_oxydex_o2ring_gap.csv
!uploads/synthetic_oxydex_o2ring_longnight.csv
!uploads/synthetic_oxydex_o2ring_longcycle.csv
!uploads/synthetic_pulsedex_rr.txt
!uploads/synthetic_hrvdex_welltory.csv
!uploads/synthetic_glucodex_lingo.csv
!uploads/synthetic_glucodex_lingo_gap.csv
!uploads/synthetic_ppgdex_verity.txt
# The FRAGMENTED Verity twin (INTEGRATOR-GAP-AWARE-OVERLAP-FOLLOWUPS §2.2). The clean twin above is
# contiguous, so PpgDex's coverage() returns null on it and nothing COMMITTED exercised the emitter —
# the gap derivation was gated only by inputs hand-built inside the test, which can drift with it.
# Same 40 s stream with two arm-off holes cut out, so diffing the two shows the gap and nothing else.
!uploads/synthetic_ppgdex_verity_gapped.txt
# O2Ring FINGER twin (PPGDEX-O2RING-FINGER-SITE §6) — single optical column, 8-bit, ~125.7 Hz,
# carrying BOTH sentinel classes. Closed-form, no recording of any person.
!uploads/synthetic_ppgdex_o2ring_finger.txt
!uploads/synthetic_ecgdex_h10.txt
# FRAGMENTED H10 twin (INTEGRATOR-GAP-AWARE-OVERLAP part 2) — three recorded segments inside a 60 s
# envelope. The only committed input in the repo that exercises `recording.coverage`; every other
# equiv fixture is gapless, which is precisely why the gap-aware defect shipped green.
!uploads/synthetic_ecgdex_h10_gapped.txt
!uploads/synthetic_oxydex_golden.node-export.json
!uploads/synthetic_pulsedex_golden.node-export.json
!uploads/synthetic_hrvdex_golden.node-export.json
!uploads/synthetic_glucodex_golden.node-export.json
!uploads/synthetic_glucodex_gap_golden.node-export.json
!uploads/synthetic_ppgdex_golden.node-export.json
# …and its FRAGMENTED sibling (INTEGRATOR-GAP-AWARE-OVERLAP-FOLLOWUPS §2.2) — the golden for the
# two-hole Verity twin, so recording.coverage has a committed equiv/GATE-C leg like ECGDex's.
!uploads/synthetic_ppgdex_gapped_golden.node-export.json
!uploads/synthetic_ppgdex_rich_golden.node-export.json
# …and the O2Ring FINGER golden (O2RING-ADAPTIVE-TIMEBASE Stage 3b) — pins PpgDex's device-crystal
# DEFAULT (the crystal path had no committed leg; every other PpgDex golden is a Verity).
!uploads/synthetic_ppgdex_o2ring_finger_golden.node-export.json
!uploads/synthetic_ecgdex_golden.node-export.json
!uploads/synthetic_ecgdex_gapped_golden.node-export.json
!uploads/synthetic_ecgdex_rich_golden.node-export.json
!uploads/synthetic_motiondex_acc.txt
!uploads/synthetic_motiondex_golden.node-export.json
!uploads/integrator_tch_golden.node-export.json

# Trio TCH corpus — DERIVED per-night node-exports (tools/trio-batch.mjs), the §6 input contract
# for `tools/tch-multinight.mjs --dir`. Tracked so a fresh clone reproduces the real σ distribution
# (INTEGRATOR-THREE-CORNERED-HAT-FOLLOWUPS-III §1: "commit ≥~5 more nights' three node-export JSONs").
# SAFE TO TRACK, and this is the one place a scoped un-ignore is warranted rather than the exact-name
# rule above: uploads/trio/ is a PURPOSE-BUILT directory whose only contents are generated by
# trio-batch, which runs every export through the shared dexScrubExport (no device/serial/model, no
# input filenames or hashes) and emits 5-min summary epochs — NO raw signal samples. The exact-name
# discipline above exists to stop a blanket glob sweeping the unrelated real-recording exports that
# live loose in uploads/ root; a dedicated subdir has no such neighbours to sweep.
!uploads/trio/
!uploads/trio/**

# Locally-attached raw capture folder (Polar Sensor Logger exports)
Ecg nightly/

# If you want to ship the synthetic (non-personal) corpus only, un-ignore it:
# !uploads/synthetic/

# ── Throwaway / generated ───────────────────────────────────────────────────
screenshots/
screens/
derive-bundle/
*.thumbnail
.thumbnail

# ── OS / editor cruft ───────────────────────────────────────────────────────
.DS_Store
Thumbs.db
desktop.ini
*.swp
*~
.vscode/
.idea/

# ── Node (none committed today, but guard against it) ───────────────────────
# NO trailing slash, deliberately. `node_modules/` is directory-only and does NOT match a
# SYMLINK — and symlinking the root's node_modules into a worktree is the documented way to
# work around a fresh worktree having none. The link then reads as untracked, so
# `tools/wt-done.mjs` REFUSES to clean that worktree up ("may be someone's only copy"),
# which is how orphaned trees accumulate — the failure §5b of CLAUDE.md measures in GB.
node_modules
npm-debug.log*

# ── Raw biosignal recordings outside uploads/ (June 2026 miss) ──────────────
ppg-nights/
analysis/*realdata*
Ecg nightly/

# ── AppleDouble / diagnostic / scratch ──────────────────────────────────────
._*
_diag/
scraps/
_texturetest2.html
_texworker.js

# ── Python virtualenvs / build junk ─────────────────────────────────────────
# The suite ships NO Python runtime: the GPU fast lane is WebGPU (sensor-trio-gpu.js),
# precisely so nothing here needs torch/ROCm and it still runs on Windows. Any venv is a
# LOCAL analysis scratch (e.g. an ad-hoc PyTorch/ROCm env) and must never reach the repo —
# they are multi-GB and machine-specific. NOTE: `*.py` is deliberately NOT ignored, because
# capture-host/*.py (the Polar capture host) is tracked and authored.
# BOTH FORMS, and the slashless one is the load-bearing addition. A trailing slash matches
# DIRECTORIES ONLY, so `.venv/` does NOT ignore a SYMLINK named `.venv` — which is exactly what you
# create when running the capture-host tests from a git worktree (`ln -s ../../Tepna/capture-host/
# .venv`), the workflow CLAUDE.md §1 tells you to use. It then shows up as untracked, and a blanket
# stage sweeps it into somebody's commit — the failure §2 spends a page warning about.
venv/
.venv/
venv
.venv
env/
**/venv/
**/.venv/
**/venv
**/.venv
__pycache__/
*.py[cod]
*.egg-info/
# large local model / tensor artefacts, if any analysis ever produces them
*.pt
*.pth
*.safetensors

# pytest-cov writes .coverage into the CWD, so a run from the repo ROOT lands it here rather than under
# capture-host/ (whose own .gitignore already covers it). Untracked-but-not-ignored is the worst state:
# it shows up in every `git status` in a tree several sessions share, where the rule is "files you don't
# recognize, leave them" — so real in-flight work gets harder to spot amid recurring build cruft.
.coverage
.coverage.*
htmlcov/
.mutation-crawl/

# Mutation SWEEPS — the survivor inventory tools/mutation-worklist.mjs ranks its queue from.
# A measurement, not a source, so it is not committed; but it must survive a REBOOT, which is why
# it lives here and not in /tmp. Eight /tmp paths were wiped by a restart on 2026-08-14 and the
# empty queue then read as a FINISHED one (0/0 = NaN% beside "target 99%").
.mutation-sweeps/

# c8 / coverage output — generated, never committed
coverage/
.nyc_output/

# mutation sweep journals — resumable-run state, regenerated, never committed
.mutate-journal/

# The §P2 fix lane's work queue: check.sh rewrites it on every run (PYTHON-TYPES-AND-FORMAT §P2).
# Machine-local and regenerated, never a source of truth — committing it would ship one box's
# error list as if it were the repo's.
capture-host/.mypy-latest.txt