feat: add SessionBank checkpoint replay telemetry

[PhilipJohnBasile]

Summary

This is a stacked draft PR on #3. It adds a default-off, bounded, content-free SessionBank overlap trace and a CPU-only offline checkpoint-placement replay/falsifier.

• records request-start-ordered RAM overlap candidates with exact-once terminal outcomes
• adds an exact irregular checkpoint-placement planner and chronological 70/30 replay
• keeps candidate identity, bank epochs, committed-MTP history, quiescence, and provenance fail-closed
• adds CLI and authenticated admin snapshot/clear endpoints
• keeps disabled tracing inert before prompt copying, scope allocation, or replay-only capture metadata
• separately hardens unknown put_snapshot recurrence provenance to recurrent/fail-closed

Contract

This does not activate a runtime placement policy and cannot emit a deployable GO. runtime_go remains false and the CLI exits nonzero even when the CPU counterfactual clears its numeric threshold.

The two replay arms are deliberately limited:

• observed_fitted may select successfully captured snapshots whose payloads were later discarded, so it is counterfactual-only.
• future_known_upper uses event-specific future knowledge and is only a no-go oracle.

Only complete, quiescent traces with no drops, pending probes, or candidate truncation are eligible for favorable evidence. Request ordering is server probe/admission chronology, not client-observed HTTP start time. The trace excludes prompt/token/session/model content, but includes prompt lengths, bounded duration counters, epochs, and process-local ordinals.

Runtime promotion still requires implementable retention/rematerialization, exact continuation, measured M5 cost, quality, regression, and physical-memory gates.

Verification

Exact reviewed commit: cce784be70da96320e05db2565581bcdd55e6000

• 636 passed, 1 intentional skip (CPU only)
• Ruff lint passed for all changed Python
• all new Python files are Ruff-format clean
• py_compile passed
• git diff --check passed
• independent Sol frozen-diff review: APPROVE, no P0-P3 findings

No GPU/model load, M5 timing, quality, physical-memory, or production-speed claim is included.

[PhilipJohnBasile]

Publication receipt for cce784be70da96320e05db2565581bcdd55e6000:

• local CPU gate: 636 passed, 1 intentional skip
• Ruff lint, new-file format, py_compile, and diff checks: passed
• independent frozen-diff Sol review: APPROVE, no P0-P3 findings
• non-PyPI release workflow: passed
  • package build: passed
  • Twine check: passed
  • fresh-venv smoke: passed
  • artifact upload: passed
  • PyPI publication: skipped by design

Workflow: https://github.com/PhilipJohnBasile/MTPLX/actions/runs/31515511225

[PhilipJohnBasile]

Final merged-state receipt:

• feat: add exact sparse checkpoint placement planner #3 merged first at 5df9af9cc87ac657bdce88f42d0a8f850d4bd29e
• feat: add SessionBank checkpoint replay telemetry #4 retargeted to the integration baseline and merged at 1d1a02662205a29f78a6f3d119300334632db40a
• the remote integration branch points to that exact merge commit
• non-PyPI release validation checked out 1d1a02662205a29f78a6f3d119300334632db40a and passed package build, Twine validation, fresh-venv smoke, and artifact upload
• PyPI publication remained disabled

Final workflow: https://github.com/PhilipJohnBasile/MTPLX/actions/runs/31515841020