LI pre-pulse auto calibration (prepulse calibration_source)

[SoichiroYamane]

Purpose

Enable JointHarmonicGls without prebuilt calibration artifacts: derive the per-channel noise model once per LI run from pulse.background_before, with digest provenance. No manual model files or hashes.

Scope

• New calibration_source field on the joint GLS config: artifact (default, current behavior) or prepulse
• prepulse: derive once per LI run (pmoke li standalone or inside analyze), per channel, from background_before; record prepulse_calibration_digest
• Validation: prepulse requires empty calibrations (ambiguous mix is an error); artifact per-channel rule unchanged; boxcar_legacy unchanged
• Mirror the validation change in pmoke-config-core; bilingual reference updates via docs-export only

Non-goals

Per-window or per-shot adaptive refit; background_before redefinition; artifact deprecation or forced migration; boxcar changes; schema version bump; failure_policy extension; live hardware operation.

Acceptance criteria

• prepulse with empty calibrations starts on both analyze and standalone li and yields LI output
• Deterministic reruns: same digest and equivalent LI output
• Named-error abort (no fallback) on empty, adequacy-failing, or non-finite pre-pulse intervals
• artifact mode unchanged (missing entry still blocks); prepulse with non-empty calibrations is an error
• noise_mode=identity with prepulse still derives and records; boxcar output unchanged
• config-core parity with native validation; bilingual refs with zero generated drift

Compatibility impact

Existing configs behave identically (default artifact). One additive provenance digest field. New configs are unreadable by old binaries (deny_unknown_fields, one direction only).

Security impact

Recorded-data only, synthetic fixtures for tests, no external I/O. No private data in the repo or PRs.

Validation plan

Nix shell: fmt, clippy -D warnings, affected unit and integration tests, docs-export zero drift, EN/JA parity. Review 1 (design/behavior) and Review 2 (staged diff) per repo gates. Merge only on green exact head with explicit authorization; close this issue explicitly afterwards.

[SoichiroYamane]

Engineering Review 1 (design/behavior, PR #261 head 2773cc0, complete working tree):

• Boundary: schema additive only (v7 kept, artifact default); no boxcar/background_before/artifact-mode/failure_policy changes; no dep/lockfile/CI changes; no live hardware. Rollback is config-only (calibration_source back to artifact).
• Derivation point: li_process_joint only, shared by pmoke li and analyze via execute_lockin. One derivation per LI run before estimation, per channel, from pulse.background_before with the sensor-stage inclusive time convention. No per-window refit.
• Digest (FR-04 layout fixed in src/lockin/prepulse.rs): domain tag + start/end bits + sample count + dt bits + acquisition digest (manifest sha256, absent sentinel only when no manifest) + fit_harmonics/noise_mode/envelope_degree; seed fixed 0, no RNG.
• Gates reused, not reinvented: plan_blocks (min 2 blocks / 1 interval, 8-target adaptive block length), calibrate-equivalent nuisance/variance/correlation builders, SCS adequacy split-half, build_artifact validation, shared artifact-to-model reduction (refactor verified message-identical via git diff -w; all 74 lockin tests pass).
• FR-07 reading: derivation runs for every noise_mode; only the correlation table is scoped to modes that need it (mirrors the loader reduction). Recorded as the one interpretation in the handoff.
• Short-interval policy (2 cycles/block floor, 8-bin qualification-scale recipe) documented with rationale in code; failure modes are named errors, never fallbacks.
• Provenance: additive prepulse_calibration_digest only in prepulse runs; artifact/boxcar manifests byte-identical (skip_serializing_if None). Derived bindings carry path prepulse:chN + artifact-bytes hash, no retained files.

[SoichiroYamane]

Engineering Review 2 (staged diff, PR #261 head 2773cc0):

• git diff --cached --check clean; stat: 24 files, +1590/-161, exactly the slice (config schema/validation x2, derivation module + tests, entry wiring, provenance, compare leg carry-through, bench/test literal updates, docs.rs + xtask schema, generated EN/JA refs + schema JSON). No lockfile/CI/workflow/version changes; generated files only via docs-export (re-ran: zero follow-on drift).
• Evidence: fmt clean; clippy -D warnings clean (pmoke, analysis-core, config-core, xtask); pmoke lib 591 passed release, bins+integration (4+4+3) release green; lockin 74 + config 124 + compare 11 debug green (nix, Python available); analysis-core/config-core/web-wasm/prologix release green; scripts python 52 OK; pytools lane N/A (dir absent pre-existing); xtask 3 green; secret-pattern grep over diff clean (gitleaks gate runs in CI).
• NFR-03: derivation 5.96ms vs LI run 986ms = 0.6% release (3.0% debug) on the AT-01 fixture; under the provisional +10%, no Architecture return needed.
• AT mapping: AT-01 e2e run_li test; AT-02 rerun-agreement test; AT-03 empty/short/adequacy/non-finite unit tests; AT-04 existing artifact tests untouched+green; AT-05 native+core mix-rejection tests; AT-06 identity-mode e2e (derives + records); AT-07 existing boxcar tests green; AT-08 core parity test; AT-09 docs-export zero-drift.
• Residual: analyze-CLI-level coverage via the shared entry point (no separate fetched-waveform CLI fixture); FR-07 scoping interpretation; additive provenance keys. Do NOT merge / do NOT close LI pre-pulse auto calibration (prepulse calibration_source) #258 without separate explicit authorization.