Preserve forked native conditioning in multi-history tokenization

[bradhilton]

Complete recorded exchanges can still fail tokenization when a later request retains or canonicalizes an earlier response and its native prompt is no longer an extension of that response's full prompt/output sequence. This change lets the existing multi_history=True and internal exchange-training paths emit adjacent native streams, each preserving the original conditioned tokens, logprobs and sampled ownership. Single-history APIs continue to refuse incompatible linear representations; no new option is introduced.

Stacked on #977. Ordinary native successes remain first. Splitting requires complete unchanged Chat provenance, exact canonical request scopes, full native source-span occupancy and STOP proof. Request-context roles retain the existing rendering proof. Final callback-free snapshots prevent a later renderer or STOP encoder from changing already-certified source/context evidence. Explicit render overrides and incomplete/edited authority do not authorize splitting.

Validation:

• Canonical CI passes at d1bb8e90e: 1,264 lightweight tests and 2,330 unit tests, plus lint, formatting, types and lockfile checks. Independent source and retained-result reviews passed; the test-only warning-isolation and nonzero-gradient follow-ups also have affected reviews.

• Public regressions independently compare first ownership before finite filtering, including NaN/overflow-first duplicates; they exercise nonzero trajectory advantages, shared weights, causal packing, copied context, strict refusals and callback mutation. Actual ART CISPO/PPO losses and gradients of a tiny shared-parameter CPU predictor match an independent original-completion analytic oracle, including clipped and active contributions of both advantage signs and physical prefix sharing.

• One bounded offline comparison of the retained batch208 trajectory: unchanged Reuse recorded tokens by default and preserve literal assistant content #977 refuses; this change preserves all 31 source encounters/30 unique sources in three histories, with 4,139 first-owned finite loss terms and 29 sampled STOPs. Public process-worker and internal preprocessing arrays agree; original input/options remain unchanged. Both single-linear APIs still refuse. The captured oracle independently derives inventory and first ownership while sharing unchanged STOP/source-key primitives. No model, provider or GPU run was performed.

• The exact production tokenizer in this PR (92311689) was reapplied as a one-file patch to the selected current ART main 80015fae receiving tree. All 60 affected receiving cases pass. A new bounded offline replay again verifies the previously refused batch208 input (31 encounters / 30 sources, 4,139 finite selected terms and 29 sampled STOPs), while a previously successful real trajectory spanning stop, length and tool-call responses preserves its entire result and all four ownership masks exactly. Original inputs, options and source bindings remain unchanged; both local process families are closed. This is CPU token/objective coverage, not a native optimizer-update or fresh learning claim.

The captured normalization check uses a constructed one-trajectory group, whose centered advantage is zero; it is not original-group gradient equivalence. The separate public gradient regression is not transformer-weight or GPU numerical equivalence. Earlier responses outside a run become that request's context, so this does not claim universal OUTPUT/SFT layout parity with every formerly successful rendered-copy path. Extra streams can repeat context; no speed or GPU-memory claim is made. Current experiments and frozen pins are unchanged.