Reject unsupported selective recompute in TrainerRank planning

[bradhilton]

A selective-recompute runtime can pass TrainerRank's activation-memory admission check and then OOM on its first forward: the current static estimate is calibrated for full recompute. Reject recompute_granularity="selective" during TrainerRank construction with TrainerRankRuntimeSupportError and an actionable hint to restore full/uniform/one-layer recompute.

This is the first step for #913. Draft follow-up #915 models retained activations and removes the temporary selective-recompute refusal; it is stacked on this PR and awaits native GPU calibration. The provider and other Megatron entry points keep their existing behavior.

Validation:

• All 276 tests across topology, TrainerRank validation, and split admission passed (269 initially; seven multiprocessing cases passed after correcting the stdin-based test launcher).
• The selective-recompute rejection tests pass at TP 1/2/4 and all three fail against the original constructor. Full/disabled recompute acceptance is covered as well.
• Changed-file Ruff lint, formatting, type checks, and uv lock --check pass.
• Repository-wide type checking reports 23 existing diagnostics in untouched files.

[bradhilton]

Review text removed on 2026-09-17; the consolidated review record is in the summary comment on this PR.

[bradhilton]

Review text removed on 2026-09-17; the consolidated review record is in the summary comment on this PR.

[bradhilton]

Consolidated review record (automated agent review traffic removed 2026-09-17)

Change. Rejects recompute_granularity="selective" at TrainerRank construction with TrainerRankRuntimeSupportError and a hint to restore full/uniform/one-layer recompute: the static activation-memory admission estimate is calibrated for full recompute, so selective runtimes passed admission and OOMed on the first forward (#913). Other Megatron entry points are unchanged.

Review. Minsky and McCarthy posted source-correctness CLEAR at 8aa6d4f; Shannon posted LGTM. Verified: topology tests pass (14/14, new selective-refusal cells at TP 1/2/4), the base fails exactly the three selective cases as a negative control, 72 further constructor controls pass, Ruff/format/ty clean. Hosted quality checks and the 2x H200 job were pending at review time.

Deferred / follow-ups.

• Model retained activations and allow selective recompute in TrainerRank #915 (stacked draft) prices retained activations and removes this temporary refusal; not covered by these reviews.
• Memory planner ignores recompute_granularity: selective recompute admitted a 32k-token group at 5.9 GB predicted, then OOM'd at >107 GB (TP 2/4, Qwen3.8-27B) #913 remains only partially resolved: None/disabled recompute is still accepted with the same uncalibrated estimate.
• Non-blocking hint gaps: EP overlap normalizes full recompute to disabled; none is not a workaround; TrainerRank docstring omits the new refusal.