P1: science corrections in results and methods (UQ-05, DET-01, SCI-05, S-RE.2-4, SCI-06/INV-01, ...)

[mdenolle]

Package P1 of the iteration-3 plan: corrections that change what the paper says about its results. Owner: agent; the author adjudicates the two decisions linked below. Size: days. Branch after P0 merges. Depends on #38. Related: #47, #50.

Closes or advances ledger entries: S-RE.2, S-RE.3, S-RE.4, S-FD.3 (= S-RE.1, S-ME.N2), UQ-05 (multiverse remnant), DET-01, SCI-05, UQ-03, UQ-04, SCI-06, INV-01, SCI-04, SCI-03, SCI-09, SCI-02, S-ME.N4, S-DI.R2 to R6.

• Multiverse on the daily grid (UQ-05, S-RP R5, S-RE.2). src/codameter/deviations.py multiverse() still decimates CCFs to every third day before stacking (lines ~345-346), so its 10-day stack spans 28 days. Stack daily CCFs and decimate only the output, as uq_bayes.run_processing_ensemble does; rerun demo_11 (~15 min); report n = 96 valid pipelines, name the failing MWCS x 4-14 s window combination and its cause, count it in the panel annotation and the Sobol basis; update Section 4 and Fig 13 from the new sidecar. The Sobol ranking may move; report the new one.
• Step-error metric (S-RE.3). fig_stack_coherence._step_bias takes a minimum over 120 post-event epochs; negative means the drop is over-estimated, and the -5.4 percent value at 1-day stacks is extreme-value bias. Define the metric in the Fig 9 caption; replace the minimum by a robust drop estimate (post-window median or a fitted step; decision in Author decision: replacement for the stacking step-error metric (blocks P1) #50); regenerate demo_18; rewrite the Section 3.6 direction statement from the new sidecar.
• Branch rule (S-RE.4, SCI-09, qmd 865-876). Rewrite as a recommendation; state that the present pipeline fits both branches jointly (branch="both", synthetic_demo.py ~344-351) and that C_d has no branch term; delete "preferring the branch of greatest change is a researcher's judgement". Per-branch measurement is deferred and said so.
• Reference datum per experiment (S-FD.3, Table 2 rows 313-320). Split the row: noise-free generating reference for the per-choice figures (best-case numbers), first-60-percent stack for OAT, multiverse and the Bayes ensemble (deviations.py:182), the five named schemes for Section 3.5. Add the estimator defaults: MWCS/WCC 6 s sub-windows, 3 s step; DTW gamma 0.3, max lag 0.8 s.
• DET-01. Reject or label inversion-reference combinations that ignore estimator and stack (deviations.py ~214-217); apply the coherence gate consistently across the reference axis or state the asymmetry (223-225); report common-support RMS with availability for the OAT sweep; add paired-seed replicates (three seeds) with a spread to Fig 12; regenerate demo_10.
• SCI-05 residual and S-ME.N4. State hyperpriors (InvGamma(2, 1e-8) for tau^2 and s^2; Gamma(2, 1e-10) for lambda), chain settings (one chain, 1200 sweeps, 400 burn-in, thin 2), seeds and the MIN_COHERENCE = 0.5 rule in the Bayes section. Add to codameter.calibration a split-member coverage check (fit s on half the members, score the other half) and a two-chain split R-hat on tau^2, s^2, lambda; rerun the clean scenario with them and report. Say in text and in the calibration-table caption that a realisation redraws the noise on one fixed coda (S-DI.R2).
• UQ-03/04 residual. Add a duplicate-configuration invariance test and a script that whitens member residuals with C_d and reports their autocorrelation (a check of R); state the result; keep the working-likelihood wording.
• Depth section (SCI-06, INV-01). Delete or make explicitly future the sentence promising a density-field "documented extension (Section discussion)" (qmd ~1251-1253); reword "The width of C_m(z) is the deliverable" and "inherits the temporal correlation and common-mode structure of C_d" (1232-1234) to "inherits whatever correlation the supplied per-band covariance carries"; reconcile "in development" (1200) with "implemented in codameter" (1230) as "implemented, not evaluated here"; fix the uq_measurement.py:33-35 and uq_depth.py:182-185 docstrings that still promise a closed loop.
• Estimator prescriptions (SCI-04, S-CD.2). Make Table 4 line 1017 agree with Results 345-348; state the Yuan et al. (2021) regime once Author input: confirm Yuan et al. (2021) Table B3 estimator ranking (blocks P1, P4) #47 confirms Table B3.
• Trailing reference (SCI-03). Table 3 row 395-396 and the "Scale of effect" sentence 704-707 to name the ablation.
• Synthetic framework (SCI-02). Reconcile "exact single-scattering solution" (191) with "multiply-scattered diffuse coda" (209); qualify "every departure ... is an artefact of the processing" (228-229) to exclude additive noise; call the envelope RT-inspired.
• Discussion (S-DI.R3 to R6). State that both bands see one imposed dv/v on the synthetic; name the comparator for "more than the measurement noise" or drop it; "committed as produced by the cloud run".

Acceptance: every changed number traces to a regenerated sidecar; the calibration table regenerates from archived runs; tests pass.

---

Working agreement (agents and humans). One PR per issue, branch iter3/<package> off docs/sign-convention-manuscript; name the ledger IDs in the PR title; request Copilot review before approval. Do not edit review/REVISION_PLAN_ITER3.md except to tick boxes and append a dated log line naming this issue. Do not hand-edit statuses in reviews/codameter-gji.review.json; the reviewer reconciles them in P7. Tests and pre-commit hooks must pass. Before requesting review, rebuild the PDF (cd paper && ../.pixi/envs/dev/bin/python build.py --no-survey) and check its text for ??, ,%, << and [Funding. Evidence for every finding: reviews/iter2/block_S-*.md; full report reviews/codameter-gji.iter2.report.md; plan review/REVISION_PLAN_ITER3.md.

[mdenolle]

Executed in PR #52 (branch iter3/P1). Two locked calibration reruns (shared_source, clock_drift) are still running and will be pushed to the same branch; see the PR body.