daily-fix: map fit-space parity assert (issue1739 natpv)

Overview / Motivation

Auto-filed by the /daily 2026-07-31 problem sweep (CONSOLIDATED M11; miner-1:P7). Source: session 55419495 (#1739) — the natural-PV pipeline applied a WHITENED-fit map to RAW activations; the resulting sycophancy headline (0.577) was quoted to Thomas and committed, and was corrected to 0.486 ~21 h later after a teammate build round found the input-space mismatch. The qualitative claim survived; the number changed. Experiment-code fix (wf_fix: false — no workflow-surface change).

Goal

Add an input-space parity check to the #1739 map-projection path: record the fit-space (whitened/raw, with the whitening params fingerprint) in the map payload and assert it against the inputs at apply/load time, so a whitened-fit map can never silently score raw activations again.

Workflow gap

• Bug observed: scripts/issue1739_natpv.py projected RAW activations through map payloads that were FIT in the whitened main-grid space (arms.py: z_ctx whitened, U-pool shrinkage whitening fit fresh in-CLI and never persisted), producing a wrong headline number that rode the interim doc/status for most of a day.
• Why it is a gap (experiment code): Nothing in the map payload ↔ consumer seam declares or checks the input space. The post-incident state at compose time: the natpv cube metadata now DISCLOSES its space ("space": "RAW activation space (NOT the whitened main-grid space)", scripts/issue1739_natpv.py:557), and a DIAGNOSTIC script exists (scripts/issue1739_map963k_applycheck.py — applies "the payload's own whitening" vs the hypothesised skipped-whitening form and reports a whitened_input_norm_mean >> sqrt(dim) out-of-distribution signal), but no durable ASSERT exists anywhere in the apply/load path — a repeat of the same mismatch would again fail silently.
• Confidence (emitter): medium-high (mismatch and diagnostic script probed; the missing-assert absence probed)
• verified-at-filing: grep -rn "fit_space\|space_parity\|input-space parity" scripts/issue1739_*.py src/explore_persona_space/analysis/leakage_predictor.py → 0 hits (absence claim: no parity assert exists — the 0-hit result is the evidence). grep -n "whiten" scripts/issue1739_natpv.py → 7 hits, incl. the raw-space disclosure at line 557 and the docstring lines 41-46 describing the whitened-fit/raw-apply seam (context read: disclosure only, no check). scripts/issue1739_map963k_applycheck.py header lines 3-23 confirm the diagnostic-only status. git log --format='%h %ad %s' --since='4 days ago' -- scripts/issue1739_natpv.py → 2 commits, latest b9920c4a06 2026-07-30 19:24 -0700 (Hub-retry waiver / load_dotenv / stage_hub_prefix — unrelated); no parity-assert fix landed on this checkout at compose time. unverified hypothesis — verify at plan time: the miner's "correction committed 2026-08-01" — no natpv.py commit after 2026-07-30 19:24 is visible here; the 0.577→0.486 numeric correction likely landed in a writeup/figures commit, not the pipeline, so the pipeline-side assert remains open.

Proposed change (candidate diff sketch — refine in planning)

Map FIT side (src/explore_persona_space/experiments/issue_1739/fits.py or
wherever the map payload .npz is written):
+ payload meta gains: fit_space: "whitened" | "raw",
+   whitening_fingerprint: sha256 of (mu, sd/shrinkage params) when whitened,
+   train_input_norm_mean (for the OOD cross-check).

Map APPLY side (scripts/issue1739_natpv.py map-load path, ~line 523
`mp = maps[kind]`, + any other consumer):
+ def assert_input_space(map_meta, X):
+     if map_meta["fit_space"] == "whitened": require the caller to pass the
+       whitening transform matching whitening_fingerprint (fail loud if absent);
+     cross-check: np.linalg.norm(X, axis=1).mean() within a tolerance band of
+       train_input_norm_mean (the applycheck OOD signal, made a hard gate).

(Reuse the norm-based signal issue1739_map963k_applycheck.py already computes; the assert is the durable form of that diagnostic.)

Scope / surfaces

• Primary target: scripts/issue1739_natpv.py (map-apply consumer) + the map-payload writer in src/explore_persona_space/experiments/issue_1739/ (fits/arms — exact writer located at plan time)
• Secondary: scripts/issue1739_map963k_applycheck.py (fold its norm check into the shared assert), any sibling consumer of the same map payloads (scripts/issue1739_pvsynth_arms.py — check at plan time)
• Experiment code — NOT the workflow surface; no .claude/ edits.

Constraints / invariants

• Fail fast — the parity check raises on mismatch; no silent fallback to raw application.
• Backward compatibility: a payload WITHOUT fit_space metadata is treated as unknown → loud warning naming the #1739 incident (or refuse, planner's call), never silently assumed raw.
• Committed whitened main-grid values and the corrected 0.486 headline are unchanged — this is a guard, not a re-fit.