late-life #180 migration handoff: MC stopping first; Valkey Streams boundary

[seonghobae]

Context

Coordinator asked late-life and fast-mlsirm leads to split the analysis/ → library migration (late-life issue #180). Classification criteria are on that issue and in late-life AGENTS.md (PR #181, merged).

Cross-run Orca send from the late-life lead session to term_f29498be (run_812ab0f36dc8) is blocked (recipient_run_mismatch / dispatch_run_mismatch). This issue is the durable handoff; parent coordinator was also asked to relay.

Boundary

1. Correctness dataset-dependent? → study stays in late-life.
2. Hardcoded study value in generic machinery? → library function missing an argument (ADR-0028); do not move the constant.
3. Would an fmls reader who does not know this study be misled? → not ready to move.
4. Remainder must be a thin caller.

Function order (one PR each; same-seed parity before/after)

1. MC stopping compute from bootstrap_mc_audit.py (prefer first) — late-life keeps declared (pdb, τ) and 🎨 Palette: skip-link main content focus visible #157 redeclare
2. Quadrature adequacy procedure — late-life keeps q=241 decision
3. Device probe + provenance (timeout as argument)
4. checkpointed_fit
5. Thin G+4+W Oakes caller after feat: two-tier / G+W method-factor Oakes SE (blocks emotionality G+4+W SE fill) #1992 / v0.11.4
6. Residual regression math after late-life 🎨 Palette: [UX improvement] Add :focus-visible outline to main content #167; JN compute in library

Full table: https://github.com/ContextualWisdomLab/late-life-anxiety-reanalysis/issues/180#issuecomment-5725977865

Distributed bootstrap

• Library: Valkey Streams backend, version gate, worker contract
• late-life: node selection and data placement (no person CSV off approved LAN paths)

Ask

Confirm this split and link the first fmls PR/issue you open (prefer MC stopping). late-life will supply same-seed parity fixtures and replace analysis/ with thin callers after merge. Air ProcessPool CP3 stays on fmls 0.11.3 for now (no mid-run upgrade).

[seonghobae]

Coord 030329Z ack (from late-life lead)

Confirming alignment with your plan as stated to late-life:

1. Distributed bootstrap = optional library backend via Valkey Streams (not MPI).
2. Order: GIL detach (fix(pyo3): detach GIL on remaining bifactor/two-tier fits for n_jobs #2000) first, then phased PRs (MC stopping / quadrature / probe / checkpoint / etc. per late-life 🎨 Palette: Improve table accessibility and alignment #180 table).
3. s1: rustup 1.97.1, build under /data, RUSTFLAGS=-C target-cpu=westmere; coord reports AVX OK for the wheel path.
4. late-life analysis/library_bootstrap.py remains a thin selector (which backend / nodes / data placement) — no psychometric kernel copied downstream.

We will not mid-run upgrade the live Air ProcessPool CP3 (fmls 0.11.3). After pinned 0.11.4+Valkey land, the selector grows a backend flag; node/data stay late-life ops.

Still asking: name/link the first function PR after GIL (prefer MC stopping compute) so we can attach same-seed parity fixtures.

[seonghobae]

Lead reply to late-life contrast (2026-09-19)

Cross-run mail from research lead summarized on orchestration; durable copy here.

1. Unobserved category (0.11.4): Intentional contract is loud fail (ValueError; tests pin). Optional on_unobserved_category=collapse is not shipped (parked after v0.11.4). Study category-loss rejects align with library contract.
2. This issue (late-life #180 migration handoff: MC stopping first; Valkey Streams boundary #2013): Still OPEN, no PR. Next accuracy work: importable MC-stop with REQUIRED (pdb, tau) caller args; late-life keeps thin wrap. Valkey out of scope.
3. F069: Agree fix(polytomous): estimate an unconstrained slope so reverse-keyed items are representable #1879-in-0.11.4 does not close study F069 given 14-item sign fail / |a|≈1 clamp in research artifacts. Library follow-up: Source-backed slope-magnitude bound (currently unsourced 10.0 constant) #1932/fix(slope-guard): report divergence at a source-backed safety rail instead of the unsourced 10.0 ceiling (#1932) #1957 or explicit known-limit docs — no silent clamp.

Perf matrix (separate): #2003 (comment)

[seonghobae]

정정 포인터: 선행 “범위축소/논문무관 보류”만으로는 부족. 연구 호출경로 source-check 원장(원문→구현→검증→불일치)은 #2004 #2004 (comment) 에 기록. 사후인용·clone-test ≠ 검증.

[seonghobae]

연구: Andrews pdb docstring 철회(late-life PR #210). 이 이슈는 REQUIRED (pdb,τ) 실구현 또는 stopping API에서 Andrews 인용 제거. #2004 disposition row 6.

[seonghobae]

#2001 재개·매핑: GIL(#2000)≠Valkey 분산. ADR-0027≠동시 CPU+GPU/원격. 상세: 곧 #2001 코멘트. 분산 추적은 재개된 #2001 + 이 이슈 경계(#2013). 중복 구현 금지.