feat(megatron): support graduated precision boundaries

[gphuang]

Summary

Supports Task 9 Option 8 for Issue #220.

• Add an optional outer precision override inside Flux sensitive boundaries.
• Preserve the existing boundary as the MXFP4 exclusion region, then select outer BF16 before inner tensorwise FP8.
• Validate that outer counts are non-negative and stay within their enclosing boundary.
• Forward and log the new fields through FluxPretrainTrainer.

Defaults remain zero, so existing recipes keep their current routing.

Test plan

• 14 targeted Flux config/layer-spec tests pass in the pinned Primus image
• Exact 57-block routing proof: BF16 {0,56}, FP8 {1–3,53–55}, MXFP4 {4–52}
• Python compile and whitespace checks pass
• Option 8 model-build smoke after review and Probe 3 verdict

[gphuang]

Result: agent review blocked before runtime smoke

Blocked. The intended Option 8 recipe routes correctly, but the framework API accepts configurations that can silently collapse the graduated layout.

• get_flux_layer_spec() creates the inner/outer backends only for automatic local-MXFP4 selection. Explicit backends, TE, absent FP4, or unavailable providers silently route every block through one backend. Graduated routing must reject unsupported paths.
• tw_fp8 can inherit a blockwise/MXFP8 recipe, and bf16 can select native linears under FP16 parameters. Validate that the configured names enforce tensorwise FP8 and BF16 semantics.
• The Primus unit test checks only 57 QKV slots. Add the complete 247-slot census plus unsupported-configuration tests.

The complete base-to-head diff, surrounding config/provider paths, existing tests, and CI were reviewed. CI is green, but its Torch unit job was skipped. Rerun agent review after these findings are fixed; the PR is not ready for runtime smoke.

[gphuang]

Follow-up: re-review confirms API and routing block

Re-reviewed exact head c1e75cf across API behavior, routing, and tests. Findings are unchanged from the previous review: unsupported backend paths silently collapse graduated routing; tw_fp8 and bf16 do not enforce their stated precision semantics; and tests cover only 57 QKV slots rather than all 247 linear slots. The relevant Torch CI job was skipped.

The PR remains draft. Dependent recipe review and the 10-step smoke remain gated until these blockers are fixed.

[gphuang]

Result: agent review approved after fixes

Review found that unsupported backend and precision combinations could silently fall back to one backend, while the routing test covered only one linear per block. Fixed in 1ef2bf6f: routing now rejects unsupported backend, provider, dtype, recipe, precision-role, and layer-count states. Tests cover all 247 declared linear slots and a mixed-precision model build.

Two independent re-reviews approved the updated diff. The PR stays draft while CI and the pinned GPU test remain open.

[gphuang]

Result: reconciled graduated-routing review approved

Merged Primus main at abc46648 into the draft branch and preserved the exact Option 8 layout: BF16 {0,56}, tensorwise FP8 {1–3,53–55}, MXFP4 {4–52}.

Independent review found one post-construction fail-open path: disabling sensitive routing while retaining boundary counts silently selected homogeneous MXFP4. Commit f8cc4e16 now rejects that state at config construction and layer-spec selection, with regression coverage.

All required CI checks and static validation pass. PR 980 remains draft; the 10-step GPU smoke is still pending compute.