fix(#116): disclose the computed interval estimator (BC vs BCa) via RelaxationResult.interval_method

[marcohost33-maker]

Summary

Issue #116, option 1 only — the honesty fix, deliberately without changing any computed number. The public surface said "BCa" in at least six places, but the leave-one-out jackknife that supplies the BCa acceleration term runs only for time grids of ≤ 60 points (a latency guard: one extra GLS refit per grid point on top of bootstrap_B), while the default grid has 80 — so the default pipeline always took bca_ci's documented a = 0.0 fallback and computed a bias-corrected (BC) interval. The "a" of BCa is exactly the acceleration; the error direction is not conservative by construction.

This PR makes the run say what it computed:

• RelaxationResult.interval_method (additive, defaulted): "BCa" (jackknife ran), "BC" (a = 0 fallback), "none" (bootstrap failed or the point estimate was not finite). Default "unreported", so a deserialised pre-The advertised BCa interval is never BCa in the default pipeline: the acceleration term is silently dropped above 60 grid points #116 report stays valid and honest — for it the estimator genuinely was not recorded.
• The 60-point gate now carries its rationale as a comment at the gate, and bca_ci / the bootstrap module docstrings state the BC fallback explicitly.
• README (3 places), docs/explanation/no-single-number.md, docs/tutorials/first-diagnostic-run.md (2 places) and the U0 docstring no longer promise BCa unconditionally; U0 (fit_uncertainty) is documented as inheriting the label.
• CITATION.cff: recorded in the Pending for the next cut block as a labelling correction, per the 2026-08-09 convention — the v0.5.0 abstract is untouched.

Deliberately not included (each changes reported numbers and needs its own anchor review, per the issue): raising/dropping the gate (option 2) and estimating the acceleration from the bootstrap replicates via the empirical-influence / infinitesimal-jackknife form (option 3). #116 stays open for those.

Scope

• Bug fix (no new feature) — a public-claim/labelling defect; no numerical behaviour change
• New diagnostic or layer (please link the motivating reference)
• Performance / sparse-path / numerics improvement
• Docs / examples / packaging only — the docs half of the change
• CI / workflow / security
• Release / packaging / publish evidence

Verification

• pytest -q passes locally — 678 passed (669 at branch point + 9 new)
• tests/test_anchors.py unchanged and green (21 passed) — no computed number changes, which is asserted by the full suite passing untouched
• ruff check src tests benchmarks clean; python -m mypy src/liouscope clean (53 files)
• .github/scripts/check_claim_safety.py passed (21 public-facing markdown files)
• If touching MANIFEST_SCHEMA.json: n/a — the run-manifest contract is untouched; the new field is additive with a default on the report dataclass
• If new external dependency: n/a — none
• If touching .github/workflows/: n/a
• If touching claims/docs/release wording: this PR removes an unsupported claim (unconditional BCa) rather than adding one; CITATION.cff pending block updated per DoD item 5
• If touching branch protection / required checks: n/a

Quality contract

• No unsupported production-ready / certified / release-complete wording introduced
• Research disclaimer preserved
• False-pass and rework risk considered, not only green CI — tests/test_interval_method.py includes a discrimination test (test_acceleration_term_is_live_when_jackknife_supplied) proving the acceleration term genuinely moves the endpoints when supplied, so "BCa" cannot decay into a relabelled BC; the gate location itself is pinned at 60/61 points so the documented wording and the code cannot drift apart silently.

Reproducibility note

No numerical results change: the field is derived from control flow the pipeline already executed. Tests use the fixed amplitude-damped qubit (H = 0, jump σ₋, rate 1.0, rho_0 = |+⟩⟨+|), seed=1, bootstrap_B=10.

Conflict note (open PRs)

#115 adds other additive fields to RelaxationResult and rewrites the default-grid derivation in relaxation.py; this PR's edits there are small and additive (one field, one comment block, one assignment), so the merge either way is a trivial additive resolution. Interaction to keep in mind at merge time: once #115's gap-scaled window lands, the grid stays 80 points, so the default label remains "BC" — the two changes are semantically independent.

Linked issues

Refs #116 (option 1 of 3 — the issue stays open for options 2/3, which change reported numbers and need anchor review).

🤖 Generated with Claude Code

https://claude.ai/code/session_01Q8z6iNAyQ6ivP68b1jDeEW

---

Generated by Claude Code

[marcohost33-maker]

Status 2026-09-07 — no issue band added, and that is the deliberate outcome

This PR targets main, so a Closes #N band would work here. I checked whether one belongs and concluded it does not — recording the reasoning so the next reader does not add one by reflex.

#116 must stay open. This PR's own body says so: it implements option 1 of 3 — the honesty fix, RelaxationResult.interval_method, deliberately without changing any computed number — and leaves options 2 (raising or dropping the 60-point jackknife gate) and 3 (estimating acceleration from the bootstrap replicates) explicitly open, because each changes reported numbers and needs anchor review. The body already uses Refs #116 rather than Closes #116, which is exactly right. A Closes band here would close an issue that is two-thirds unresolved. Verified via closingIssuesReferences: this PR closes nothing, and should continue to close nothing.

State, measured:

mergeStateStatus  BEHIND      (mergeable = MERGEABLE, no conflict)
check-runs at cf522a21: 20, of which 10 are test-matrix jobs, 0 non-success
vs pr107-fix@bac5fe0d: 62 commits missing here, 1 unique to this PR

BEHIND under strict status checks, not blocked and not conflicting. Its single unique commit is the additive interval_method field.

What Marco has to decide

At 62 commits behind the #121 head, the green rollup here describes a tree that has since moved a long way. The interaction the body already flags is the live one: #115 rewrites the default-grid derivation in relaxation.py, and this PR's edits there are one field, one comment block, one assignment — additive either way, but the merge order decides who resolves it. Merging #121 (and then #127) first, and updating this branch afterwards, is the ordering that leaves the least to reconcile by hand.

Read-only measurement. No push, no merge, no band added, nothing resolved.