Experiment: add offline residency simulator

[terrizoaguimor]

Summary

This is the self-contained Phase 0 offline residency experiment requested in PR #893 review. It adds no runtime behavior, backend changes, routing changes, or prefetch.

It provides a reproducible simulator for the question raised by #884: under a fixed expert-residency byte budget, can admission/frequency policies reduce predicted critical-path felt wait without overfitting?

Included

• Strict GLM ROUTE_TRACE parsing and contiguous call/batch validation.
• First-seen batch unions plus per-row selection multiplicity.
• GLM 64-expert working-set block replay.
• LRU, LFU, synthetic pinned-LRU, SLRU, and frequency-admission policies.
• Uniform and policy-specific dynamic layer capacity allocation.
• Separate resident bytes, read bytes, and felt miss costs.
• Category-held-out decision gate: at least 10% mean improvement and no category worse than 3%.
• Layer-specific felt-cost sensitivity and aggregate PROF=1 calibration parsing.
• Complete JSON output with trace sources, specs, allocations, gates, and sensitivity.
• Deterministic synthetic fixtures demonstrating both locality wins and overfit regressions.
• 29 focused tests.

Real-trace context

The H200 pilot used 13 GLM-5.2 traces collected before this split: 8 training and 5 held-out across coding, chat, reasoning, multilingual, and long-context prompts. The offline results were:

Expert budget	Candidate	Mean held-out gain	Worst category
120 GB	Uniform frequency	+39.91%	+34.84%
120 GB	Dynamic frequency	+40.45%	+35.41%
160 GB	Uniform frequency	+53.35%	+49.92%
160 GB	Dynamic frequency	+54.50%	+50.66%
200 GB	Uniform frequency	+65.31%	+62.64%
200 GB	Dynamic frequency	+66.29%	+63.07%

These are simulator predictions, not runtime speedup claims. The pilot also showed that hit rate alone is insufficient: RAM_GB=120, PIPE=0 measured 1.66 tok/s at 57.2% hit, while RAM_GB=120, PIPE=1 measured 2.00 tok/s at the same hit rate.

Related issue

This is the simulator half of #893 and supports #884. It is intentionally independent of the placement/backend issues in #887, #892, #813, #585, #759, #766, #767, #687, #856, #885, #848, #662, and #783.

Verification

• Full Python suite on the source branch: 407 passed, 57 skipped.
• Focused simulator tests: 29 passed.
• No C runtime files are changed by this PR.

The next step is a separate runtime policy PR only after maintainers accept the methodology and held-out gate.