Add an explicit native representation for sampled tokenization

[bradhilton]

Recorded tool-call completions can make template-based tokenization fail even when the original prompts and output tokens are complete. This adds representation="native" to art.tokenize_sampled(...), letting callers build histories directly from the recorded evidence.

Native mode verifies complete conditioning, token IDs, logprobs, source order and STOP flags. It joins compatible consecutive generations and keeps others separate. representation="rendered" remains the default, and generic art.tokenize is unchanged. This PR depends on #973.

Public regressions and full CI pass. Independent Astra and Fable source reviews found no blockers. One bounded invocation on a retained failure input also passed all 30 sources’ token, logprob, STOP and ownership checks on the current head.

This opt-in representation supports the SAMPLED first-occurrence loss contract. It does not establish numerical-training equivalence or interchangeable behavior for OUTPUT/SFT losses or arbitrary per-history weighting.

Validation details

• CI for the current head: lint, formatting, type and lock checks pass; 1,228 lightweight tests and 2,056 unit tests pass, with 34 skipped and 2 deselected.
• 53 public native cases cover source integrity, tool metadata, STOP suffixes, repeated ownership, serialization and dispatch. The combined Add explicit sampled-output STOP certification #973/Add an explicit native representation for sampled tokenization #976 focused suite passes 108 cases.
• GPT-6 Astra xhigh and Claude Fable 5.1 Extra High each returned PASS_WITH_NOTES for head 3c28a3fe9; a separate changed-source peer also passed.
• The current-head retained-input witness returned two histories/21,064 tokens, 5,384 finite sampled terms and 29 correct STOP bits, with zero missing/extra bits. Input, options and compact roundtrip were unchanged. The runner used retained model metadata and local tokenizer substitutions.
• This captured result covers native mode and the shared Add explicit sampled-output STOP certification #973 guard, not the rendered route or Prove native conditioning before preserving literal reasoning markers #974 on that input. No private payload is included.