fix(storage#782): exclude write-side teardown from §4.7.1 failover-callout gap

[FileSystemGuy]

Summary

Fixes #782. The split-mode failover-callout gap check charged the write phase's post-benchmark work — final multi-node cluster collection, metadata serialization, interpreter shutdown — against the 30-second budget by measuring from write.summary.end_time (end of the timed section). On large topologies that teardown can take tens of seconds, causing false failures on legitimate submissions whose real inter-invocation handoff is within budget.

Adds a symmetric write-side bookend to mirror the read-side fix from #714:

• _invocation.py — new INVOCATION_END + mark_invocation_end(), symmetric with the existing INVOCATION_START.
• benchmarks/base.py — calls mark_invocation_end() at the top of write_metadata(), the latest moment where the value can still land in the persisted file. Emits it as metadata['invocation_end_time'].
• Submission checker cache_flush_validation — prefers write.metadata.invocation_end_time as the write-side origin, falling back to write.summary.end_time for older results dirs (pre-fix behavior). Emits an INFO line per pair with the chosen origin so the measurement is transparent regardless of pass / warn / fail.

Now the two bookends are genuinely symmetric: whatever falls outside INVOCATION_START / INVOCATION_END is excluded from the gap on both sides — framework startup on the read side, and post-benchmark cluster collection, rank-0 teardown, JSON serialization, and interpreter shutdown on the write side.

Worked example (from #782)

Quantity	Value
write.summary.end_time (timed section ends)	2026-07-13T06:36:25
write.metadata.invocation_end_time (nodes released, rank 0 exiting)	2026-07-13T06:37:03
read.metadata.invocation_start_time	2026-07-13T06:37:29
gap charged before fix (read_start − summary_end)	64.4s → FAIL
gap charged after fix (read_start − invocation_end)	~27s → PASS

Test plan

• mlpstorage_py/tests — 850 passed (includes 3 new §4.7.1 tests)
• tests/ — 2907 passed, 1 skipped, 13 deselected
• vdb_benchmark/tests — 174 passed
• kv_cache_benchmark/tests — 238 passed
• Reviewer sanity-check on a real split-mode submission whose failure motivated the issue

New tests

• test_split_mode_write_teardown_excluded_from_gap — issue Validation §4.7.1 checkpointCacheFlushValidation unfairly charges the write phase's final multi-node collection to the failover-callout budget #782 scenario: 45s teardown + 25s real gap ⇒ PASS.
• test_split_mode_missing_invocation_end_time_falls_back_to_summary_end — pre-Validation §4.7.1 checkpointCacheFlushValidation unfairly charges the write phase's final multi-node collection to the failover-callout budget #782 results dirs still measure against summary.end_time (unchanged behavior).
• test_split_mode_info_line_emitted_on_pass — INFO transparency line is emitted on every pair.

[github-actions]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅