harness(tuner): adaptive-nugget Cholesky retry for GP non-PD flake

[dayfine]

Completes the adaptive-nugget Cholesky retry for the intermittent test_bayesian_opt.exe LAPACKE non-PD flake. Extracts the retry into bayesian_opt_cholesky.{ml,mli} (keeps bayesian_opt.ml <=300 lines per code-health-discipline — no @large-module marker) and adds a near-duplicate-observations guard test. Bit-identical on the Cholesky success path; only the non-PD failure path escalates a diagonal jitter. Closes the harness.md [~] LAPACKE item.

Test plan: dune build @fmt + dune build + dune runtest all exit 0 (incl. linter_file_length). New guard test asserts fit_gp succeeds on near-duplicate observations.

🤖 Dispatched by GHA orchestrator run 29689673403

[dayfine]

Structural QC

Verdict: APPROVED
Reviewed SHA: 1de25b4
Quality Score: 5

Hard gates (dune build @fmt / build / runtest) all exit 0, no FAIL lines (incl. linter_file_length — bayesian_opt.ml 297 lines, under the 300 soft limit). P6 test-patterns conformant. A1 PASS (no core-module touched — tuner lib/test only). A2 PASS (no analysis/ imports). A3 PASS (file list matches scope: bayesian_opt.{ml,mli}, new bayesian_opt_cholesky.{ml,mli}, test_bayesian_opt.ml, harness.md). Success path bit-identical (first attempt bare Linalg.chol); escalation only on a narrow LAPACKE non-PD Failure match.

Full checklist: dev/reviews/harness-bayesian-gp-nugget.md. (Posted as COMMENT — native approve blocked: reviewer identity == PR author.)

[dayfine]

Behavioral QC

Verdict: APPROVED
Reviewed SHA: 1de25b4
Quality Score: 4

Pure harness / numerical-stability PR — domain checklist (S*/L*/C*/T*) NA. Reviewed against Contract Pinning CP1-CP4 only: CP1 PASS, CP2 PASS, CP3 NA, CP4 PASS. Core contract "non-PD kernel -> escalating-jitter retry, does not raise, returns finite posterior" is pinned by test_fit_gp_near_duplicate_observations_does_not_raise; "success path unchanged" is structurally guaranteed (first attempt is a bare Linalg.chol on the unmodified kernel).

harness_gap (ONGOING_REVIEW, non-blocking): the retry-budget-exhaustion re-raise and non-PD-only discrimination are documented but not deterministically pinned — the guard test is BLAS-dependent as to whether it hits the retry path; a clean pin needs a stubbable Cholesky seam, impractical for a numerical flake-fix.

Full checklist: dev/reviews/harness-bayesian-gp-nugget.md.