fix(meg): make refutation identifiable, not just reachable

docs/golden-validation-recipe.md

@@ -45,7 +45,7 @@ A correct finding carries the material to **sanity-check** it, not just a verdic
 | `verdict` | `supported` | in-window peak from a reliable sample |
 | `latency_ms` | `100.0` | the planted peak |
 | `expected_window_ms` | `[80.0, 120.0]` | the auditory M100 window |
-| `search_window_ms` | in-window bounds | provenance of the measurement (not the old ±100 ms widening) |
+| `search_window_ms` | `[30.0, 235.0]` | provenance of the measurement. Floor excludes the ~15 ms artifact; ceiling clears the 215 ms refutation threshold so the claim could actually have failed |
 | `n_epochs` | `40` | above `MIN_RELIABLE_EPOCHS`, so the average is trustworthy |
 | `channel` | the peak channel | which sensor carried it |
 | `caveats` | `[]` | nothing to flag |
@@ -57,7 +57,7 @@ thin, out-of-window sample:
 
 ```python
 from matilde_plugin.engine.meg_study import build_golden_steps, MIN_RELIABLE_EPOCHS
-steps = build_golden_steps(peak_latency_ms=300.0, n_epochs=MIN_RELIABLE_EPOCHS - 1)
+steps = build_golden_steps(peak_latency_ms=225.0, n_epochs=MIN_RELIABLE_EPOCHS - 1)
 ```
 
 This does **not** return a confident `refuted`. Because the epoch count is below

docs/meg-validation-study.md

@@ -31,10 +31,93 @@ response to standard tones and validates it falls within the expected window.
 3. `epoch` — epoch around standard-tone events (the analysis must read the events
    from the data, not assume counts); checkpoint epoch metadata.
 4. `evoked` — average to the evoked response; checkpoint the evoked array artifact.
-5. `validate_finding` — measure peak latency in the M100 window; emit a finding:
-   `supported` if within the expected window (default 80–120 ms), `refuted` if
-   clearly outside, `inconclusive` if SNR too low / sample too small. Record the
-   measured latency + amplitude as evidence.
+5. `validate_finding` — measure peak latency in the M100 window; emit a finding.
+   `supported` if the peak is inside the expected window (default 80–120 ms).
+   `refuted` only if it is later than the **refutation threshold** — the later of
+   215 ms (the end of P200 territory) or `window_hi + 60 ms`. Everything else is
+   `inconclusive`, with a caveat naming *why*. Record latency + amplitude as
+   evidence.
+
+   The peak search runs from **30 ms** (artifact-exclusion floor) to **20 ms past
+   the refutation threshold**, capped at 280 ms and clamped to the epoch.
+
+### Why the verdict geometry looks like this
+
+The peak is a polarity-blind argmax (`mode="abs"`). Two failure modes pull in
+opposite directions, and both have actually happened here:
+
+- **Search too narrow → `refuted` unreachable.** Between `e079993` and
+  2026-07-30 the search was window ±10 ms while refutation required window
+  ±50 ms, so **no measurement at any configured window could return `refuted`**.
+  The prediction could not fail.
+- **Search too wide → the argmax catches something that is not the M100.** The
+  original bug was exactly this: a ±100 ms search grabbed the ~15 ms stimulus
+  artifact. The first attempt to fix the falsifiability problem (2026-07-30)
+  widened the search to a 40 ms floor while leaving an early refute threshold at
+  60 ms — reproducing the same incident at a new latency, with a *confident*
+  `refuted` where the previous code had safely said `inconclusive`.
+
+Width alone cannot resolve that. So the search is wide enough to contain the
+refutation threshold, and the classifier refuses to refute anywhere a known
+component could be the winner:
+
+```
+   30 ms         70 ms      window     170 ms    215 ms       280 ms
+ ---|-------------|---------[====]-------|---------|------------|---
+   floor       P50/Pa     supported   off-window  P200      refutable -->
+ (artifact)   ambiguous               late      ambiguous
+              -> inconclusive                -> inconclusive
+```
+
+**Early refutation is not achievable this way, and the code does not pretend
+otherwise.** Everything early enough to contradict the M100 is also P50/Pa
+territory (this dataset has a P50 around 50 ms). Establishing a genuinely absent
+M100 needs amplitude/SNR or topography, not latency — see "Still open".
+
+Three guards make the remaining verdicts honest:
+
+- **`boundary_hit`** — an argmax pinned within 1 ms of a search edge means the
+  true extremum is probably *outside* the range. Forced to `inconclusive`.
+- **`component_ambiguous`** — a peak in P50 or P200 territory is named as such
+  rather than scored.
+- **`unfalsifiable_late`** — if the clamped epoch cannot reach the refutation
+  threshold, `supported` was the only attainable answer, and it is labelled
+  untested rather than confirmed.
+
+> **Falsifiability invariant, test-enforced:** the search range must strictly
+> contain the late refutation threshold. `_SEARCH_BEYOND_THRESHOLD_MS` is the
+> margin that guarantees it. If you narrow the search to chase an artifact, raise
+> the floor — do not let the floor cross a refute threshold. Enforced by
+> `test_late_refutation_is_reachable_from_the_real_search_range`,
+> `test_search_range_contains_the_late_refutation_threshold`,
+> `test_peak_at_the_search_floor_is_not_a_confident_refutation` and
+> `test_falsifiability_holds_under_the_shipped_epoch_bounds`; the constants are
+> mutation-tested (7/7 mutants killed).
+
+**Consequences for the record.** The single historical `refuted` predates the
+artifact fix and came from measuring the 15 ms artifact — **retracted**; it was
+never evidence against the M100. Every `supported` recorded between `e079993`
+and 2026-07-30 carried **no falsification risk**. Re-running the *golden* recipe
+does not re-test anything (its synthetic backend returns a planted constant); the
+number that needs re-measuring is the real **120.0 ms** from `e079993`, which sat
+exactly on the old search's upper bound and disagrees with this dataset's
+published N100 (~95 ms) by 25 ms.
+
+### Still open
+
+- **Absence cannot be refuted.** `get_peak` always returns an argmax, so "no
+  M100" is indistinguishable from "some peak somewhere". Needs an amplitude/SNR
+  threshold.
+- **Early refutation** needs polarity or topography, per above.
+- **The band magnitudes (70 / 170 / 215 ms) are conventional, not fitted.** They
+  bracket the canonical N1 (60–150 ms) and P2 (150–250 ms) and the M100's
+  reported ~99–153 ms range across frequency and intensity (Roberts et al. 2000),
+  but they are not derived from this dataset. Refutation therefore requires a
+  large latency departure — it is a guard against gross misidentification, not a
+  sensitive test of latency shifts.
+- **`MIN_RELIABLE_EPOCHS = 8`** gates `refuted` but never `supported`. That
+  asymmetric evidentiary bar deserves its own review; 8 is also well below CAEP
+  practice.
 
 ## Implementation rules (match the repo)