mesh: amendment 1 - reclassify the v4 plan, close every post-outcome choice

The 32ab699 commit message overclaimed: records existed on disk at
commit time and log tails had exposed two endpoint lines, so 'before
outcomes exist to peek at' was false when written. Reclassified as a
mid-collection, partially outcome-blinded prospective analysis plan.

Decisions closed: interval-censored event definitions on author
observation times with an explicit ambiguous bin; no formal hypothesis
tests for v4 at all; Q1 primary = v4's five runs with v3 as historical
replication and pooling gated on an acquisition-path equivalence audit
(d1366561 -> c899a021); Q2 comparator = v4 Cell A truncated per the
stated rule, Newcombe hybrid-score 95% for the risk difference; Wilson
intervals at n=3; fixed denominators with a no-replacement missingness
policy; cell key extended with topology, sampling and gate config, and
acquisition hash; controls endpoint predeclared; immutability redefined
as raw-series-plus-acquisition, extraction re-derives from series and
never reads in-record analysis fields.

32ab699 is not rewritten; where they conflict, this amendment governs.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: Fede654

tests/mesh/experiments/post-expiry-prereg-amendment-1.md

@@ -0,0 +1,106 @@
+# Amendment 1 to the v4 analysis plan (32ab699)
+
+Timestamped by its own commit, while the batch is still running.
+`32ab699` is left untouched; where this amendment contradicts it, the
+amendment governs. External audit found the plan not yet fully
+specified — several choices could still have been made after seeing
+outcomes. Each is closed here.
+
+Batch state at this amendment: Cell A complete, Cell B in progress
+(records exist on disk, unopened beyond what is disclosed below).
+
+## 1. Status reclassified — this is not preregistration
+
+`32ab699` and its commit message called itself a preregistration made
+"before outcomes exist to peek at". **That claim was false when made**:
+at commit time Cell A's five treatment and three control records and
+two Cell B records existed on disk, and log tails during health checks
+had exposed endpoint lines ("absent at window end: True") of at most
+two runs. Git verifies the timestamp; nothing can verify the files were
+never opened.
+
+Correct designation: **a mid-collection, partially outcome-blinded
+prospective analysis plan**, with the endpoint exposure above. Valuable
+because the analysis choices predate reading any resurrection count,
+trajectory, reset delta, or late-recurrence outcome — not equivalent to
+registration before collection.
+
+Auditor transparency, carried into the record: the audit that prompted
+this amendment inspected v4 schema/provenance fields and one validity
+flag, not substantive outcomes.
+
+## 2. Event intervals and horizons, exactly
+
+Resurrection witnesses are interval-censored. For every witness at the
+author:
+
+- **Event interval** = (previous author observation `done`, current
+  author observation `done`], both in per-node observation time — never
+  the shared row timestamp `t`.
+- **Evidence within horizon H**: interval upper bound ≤ H.
+- **Definitely late (threshold L)**: interval lower bound > L.
+- **Ambiguous**: interval straddles H or L — counted in neither bin,
+  reported separately, never dropped.
+- **Horizon truncation** (Q2's 3-lifetime reanalysis): a truncated
+  series keeps, per node, exactly the observations with `done` ≤ H;
+  the analyser then runs unchanged on that subset.
+
+Lifetimes are multiples of `configured_insert_ttl` (96 s → H = 288 s
+for Q2; L = 192 s and 480 s for Q3).
+
+## 3. No formal hypothesis tests for v4
+
+The conditional "if any formal test is run" left a post-outcome choice
+open. Closed: **no formal hypothesis tests will be run on v4 data.** No
+Fisher, no Holm family, no p-values. Reporting is: raw per-run
+outcomes, Wilson 95% intervals on proportions, and one exploratory
+effect estimate (Q2's risk difference). Anything beyond that in a
+future write-up is post-hoc and must be labelled so.
+
+## 4. Q1 primary estimate is v4-only
+
+v3 was acquired by script `d1366561…`, v4 by `c899a021…`, with
+substantial changes between them; reanalysis aligns derived fields, not
+acquisition behaviour. Therefore:
+
+- **Primary Q1 estimate: the five v4 Cell A treatment runs.**
+- v3's three runs: reported separately as historical replication.
+- A pooled 8-run figure is secondary at most, and only after a
+  documented acquisition-path equivalence audit (a reviewed diff of
+  `d1366561…` → `c899a021…` showing no change to probing, sampling
+  cadence, gating, or recording of the raw series) — matching binary
+  and cell dict alone is insufficient.
+
+## 5. Remaining specifications closed
+
+- **Q2 comparator**: v4 Cell A only (the five primary runs), truncated
+  to the 3-lifetime horizon by the rule in §2. v3 is not a comparator.
+- **Interval methods**: proportions get Wilson 95% intervals; the Q2
+  risk difference gets a **95% Newcombe hybrid-score interval without
+  continuity correction**.
+- **Q3 intervals**: Wilson 95% intervals ARE reported at n=3 (0/3
+  leaves an upper bound near 56%, and that width is the message).
+  "No interval worth reporting" in 32ab699 is withdrawn.
+- **Missingness policy**: planned denominators are fixed by the
+  committed driver — treatments 5 (A) / 3 (B) / 3 (C), controls 3 (A) /
+  2 (B). Every report states planned n, valid n, invalid filenames with
+  their failure reasons, and best/worst-case bounds for each binary
+  outcome (all-invalid-runs-positive / all-negative). **Invalid or
+  timed-out runs are not re-run for v4** — selective repetition after
+  seeing which runs failed is outcome-dependent sampling.
+- **Cell key**: the full record `cell` dict plus window, arm, topology
+  (chain; `directed` flag), `sample_gap_s`,
+  `propagation_check_t_requested`, and the acquisition-script hash.
+  Batch (v3/v4) is a stratum, never silently pooled.
+- **Controls endpoint**: predeclared as `absent_at_end` by cell with a
+  Wilson 95% interval, reported under the standing label "resurrection
+  not observable (3 samples)".
+- **Immutability, corrected**: "source records stay immutable"
+  conflicted with `--reanalyse`, which rewrites derived fields in
+  place. The immutable object is the **raw series plus acquisition
+  metadata**; the extract stage hashes that portion of each record
+  separately and re-derives all statistics from the series itself, so
+  v4 aggregate results never depend on in-record `analysis` fields.
+  `--reanalyse` remains a convenience for humans reading single
+  records; it is not an input to the pipeline. No `--reanalyse` runs on
+  v4 records before extraction is complete.