fix(config): reject unsupported bifactor JMLE before fitting

[seonghobae]

Public contract defect

Protected main a7be9eb8ca0b807364173f8471fc300e3cf9350f accepts FitConfig(model="BIFAC2PLM", estimator="jmle") through FitConfig.validate(), but public fit() later raises NotImplementedError because BIFAC2PLM is supported only by the marginal/MMLE estimator. This is the same API-design failure class as #639, but it is a distinct model × estimator compatibility constraint rather than an unimplemented estimator identity.

A production configuration object that reports itself valid must not later reveal that the selected model/estimator combination is an unimplemented stub. The unsupported combination should fail at configuration/compatibility validation with a stable bounded ValueError before response preparation or fitting work.

Sequencing / writer lease

PR #640 currently owns python/fast_mlsirm/config.py while narrowing the estimator vocabulary. Do not create a competing source branch or edit the same validation surface while #640 has an active writer. After #640 stabilizes or integrates, refetch protected main and reuse its final vocabulary/error conventions.

Test-first acceptance

1. Add a realistic RED at the public configuration/fit boundary proving BIFAC2PLM + jmle is rejected by validation/compatibility preflight, not later by NotImplementedError inside fit().
2. Preserve BIFAC2PLM + mmle exactly, including existing Rust-owned bifactor likelihood/calibration semantics and scoreability boundaries.
3. Preserve all currently supported non-bifactor JMLE/MMLE model combinations; do not infer future compatibility from model names alone.
4. Prefer one explicit model-estimator compatibility table/policy with a finite stable error surface over scattered late if ...: raise NotImplementedError branches.
5. The compatibility error must be non-reflective with respect to caller data and must run before response materialization/numerical work.
6. Remove the now-unreachable late NotImplementedError or retain only a defensive assertion/internal invariant that cannot be the normal public path.
7. Add exhaustive matrix tests over every current VALID_MODELS × VALID_ESTIMATORS supported/unsupported pair so future model additions cannot silently advertise unusable combinations.
8. Complete public docstrings and exact changed-production statement/branch coverage; add authoritative changelog/doctoring and render/check the managed changelog.
9. Require one unchanged head with full Python/Rust/PyO3/package/GPU/fuzz/Security/SAST/current-head review evidence before Ready.

Non-goals

• no new bifactor JMLE estimator in this slice;
• no change to bifactor scoring/scoreability/model-selection mathematics;
• no workflow/dependency/version/release change;
• no canonical PRD/TRD/Architecture/ADR/UML/ERD branch; cross-cutting documentation belongs only in docs: establish canonical product and architecture baseline #604/docs: close canonical documentation fitness gaps without parallel authority #621.

Closure

Close only after protected main validates the current model × estimator compatibility matrix up front, BIFAC2PLM + jmle cannot reach an advertised-but-unimplemented fitting path, and supported combinations remain behaviorally unchanged.

[seonghobae]

Protected-main closure evidence: PR #702 originally implemented the configuration-time BIFAC2PLM + jmle rejection, a later Sentinel merge accidentally reverted it, and PR #713 restored that accepted contract on current protected main (c8c834edb11e19952bf7e84ee384b37f471ae563). Current FitConfig.validate() again rejects the unsupported bifactor/JMLE pairing before fitting while preserving bifactor/MMLE. The transient re-repair Draft #711 is closed superseded. Closing this issue as implemented/restored on protected main.