feat(experimental): add the offline replay trace contract and converter

[0z5a]

Incremental review: Files changed against main.

Summary

Adds the UniRL side of the optional AISimulate / Dynamo Replay integration proposed in #484:
a versioned trace contract and the workload-to-replay converters, with their CPU contract
tests. It is deliberately the half that can be built and verified without a simulator.

experimental/aisim/schema.py — the trace contract, stdlib only:

• Three separated inputs: manifest (configuration and provenance), workload (what a
  predictor may read), measured outcomes (report-only). The separation is structural, not
  conventional: no converter function accepts a measured result.
• Validation fails loudly on schema/version drift, non-finite or negative times, bool/float
  where an int belongs, missing lengths, unknown ids, self-referential or missing
  predecessors, dependency cycles and unsupported status.
• A dependency arrival requires an explicit external_delay_ms (0.0 must be written
  deliberately), and a first-turn request carries an absolute release offset instead of a
  fabricated predecessor, so a missing wait can never be read as a zero wait.
• failed / aborted / partial / retry / cancelled terminations raise instead of being
  dropped, and capture_status="incomplete" cannot produce a fidelity run.

experimental/aisim/convert.py — the converters:

• to_open_loop preserves observed arrivals verbatim (rebased to a recorded origin) and adds
  no session_id, no wait_for and no synthesised tool wait; it also refuses to run under a
  dependency contract and refuses a window that mixes both arrival kinds.
• to_dependency emits predecessor-plus-external-delay entries and keeps measured service time
  out of the predictor input; assert_no_outcome_leakage re-checks the converted rows.
• Trajectory slots and collector barriers are not modelled. The sidecar says
  slot_semantics="not_modelled" and barrier_semantics="not_modelled" rather than inventing a
  representation that would make the replay look more faithful than it is.

Deliberately absent: runner.py and requirements.txt. The runner is the only file allowed to
import a simulator API, and it should land once a pinned simulator revision is confirmed instead
of being written against a guessed interface.

Related Issue

Refs #484. This is the adapter slice of that RFC; the RFC itself stays open for the interface
and profile questions it asks maintainers.

Test Plan

Environment: jk01, 8× L20 (driver 570.86.10), torch 2.11.0+cu128, base SHA
f8d95462837df00188dab1376e19347b43996307.

Lint and guards:

ruff check experimental/aisim/                     -> All checks passed!
ruff format --check experimental/aisim/            -> already formatted
python lint/check_docstring_lines.py               -> all one line
python lint/check_experimental_boundaries.py       -> ok (no core→experimental import, no sibling import)
python lint/check_core_dependencies.py             -> ok

CPU contract harness (repo-external, not committed), 18/18 not-FAIL — 16 PASS, 2 NOT_RUN:

python evidence/harness/test_rfc_p1_cpu.py --out evidence/validation_results_rfc_cpu.json
  T01 PASS  schema/version/unit/token range rejection, incl. NaN/inf and bool-as-number
  T02 PASS  unique ids; measured results matched exactly once; duplicate/missing/unknown rejected
  T03 PASS  fan-out rows stay one request each with a reverse id→group map
  T04 PASS  observed arrivals verbatim; no session/wait injected; contract mismatch rejected
  T05 PASS  time origin recorded and offsets rebased; non-ms units rejected
  T06 PASS  measured outcomes invisible to the converter; a leaked key is detected
  T07 PASS  full-coverage claim without a profile id, and missing provenance, rejected
  T08 PASS  importing and running the adapter loads no torch/ray/vllm/sglang
  T09 PASS  incomplete capture refused instead of reported as accurate
  T10 PASS  successor arrival rides predecessor completion plus the measured external delay
  T11 PASS  0.0 preserved; unknown wait and unanchored first turn both rejected
  T12 PASS  slot/barrier semantics labelled not_modelled rather than faked
  T13 PASS  equal-arrival tie-break is a stable function of the input sequence
  T14 PASS  retry/abort/partial lifecycles refused, so no window is renamed "complete"
  T15 PASS  fidelity claim without profile/topology provenance rejected
  T16 PASS  workload JSONL round-trips byte-identically and revalidates
  T17 NOT_RUN  request-level simulator output: no simulator is installed
  T18 NOT_RUN  external-runner failure handling: no runner process exists

Capability probe that scopes this PR (P0 of the RFC, recorded in the evidence directory):
import dynamo|aisimulate|spica → ModuleNotFoundError for all three; no such package in
pip list; no AISimulate/Spica reference anywhere in the checkout; the two documented Dynamo
pages and the ai-dynamo PyPI name do resolve. Hence: contract and converters only, everything
simulator-facing NOT_RUN.

Compatibility / Risk

• New package under experimental/; no existing module imports it, no core file is touched, and
  no dependency is added (stdlib only). Normal training is unchanged and the package can be
  deleted without a trace.
• SCHEMA_VERSION (unirl.rollout.trace.v1) is stamped into every record and into the manifest;
  a field rename is a breaking change for the evidence artifacts, not for the runtime.
• The exported workload is only as trustworthy as the capture hook that will produce it. That
  hook does not exist yet, which is why every capture-dependent claim is left NOT_RUN.
• No timing model, cache model, topology model or K/V model is introduced here.

Reviewer Notes

• Duplicate-work check (2026-09-18, 63 open PRs, keyword searches for AISimulate / replay /
  simulator): no PR claims a simulator integration, and "AISimulate" returns zero results. The
  closest replay work (feat(diffusion): batch video replay timesteps #394, perf(bagel): pack FlowGRPO replay steps #384, feat(flowgrpo): add bounded-memory chunked replay backward #465) is training-side batching and is unrelated.
• Worth reading first: to_open_loop's refusal paths and assert_no_outcome_leakage — they are
  the two places that keep a predictor from being handed the answer.
• Open questions deliberately left to the RFC: which capture hook to use, which pinned Replay
  interface is first, and whether the chosen runner can emit per-request completions at all. The
  adapter reports AGGREGATE_ONLY rather than reconstructing group tails from percentiles if it
  cannot.

[zzhuoxin1508]

Same note as on #484: we currently only have H20. If the Replay / AI Configurator tables have no calibrated H20 profile, I don't think this adapter should land yet — the H20 data would need to come first, otherwise we cannot validate replay against real execution.