Include captured conditioning in sampled-source identity

[bradhilton]

When captures reuse response metadata and generated tokens but have different captured prompts, source deduplication can discard a valid sampled prefix. Include captured prompt IDs in the existing fingerprint for Chat, Completions, Responses and Messages, following each protocol's extraction precedence. Identical copies still deduplicate; unrelated choices do not change the selected source identity.

For example, captures [1] → [2] and [1, 2, 3] → [2] with identical response IDs/logprobs now retain both sampled prefixes. This changes one runtime function; public tokenization arguments, loss normalization and reconciliation policy are unchanged.

Stacked on #885; partial follow-up for #883. The documented explicit/manual changed-conditioning and renderer-repair counterexamples remain unresolved, so this PR does not close #883. Held for morning merge review; existing research drivers require separate source resealing before adoption.

Validation: independent review at exact head 9fe17e39db8ad237ae304920b158b69f2f66ebf9 passed all 484 trajectory tests and 21 additional duplication/retention controls. The 16 new tests produce 10 expected failures and six passing controls on the parent. Two real captured trajectories retain exact tensors and 7,445/11,685 selected tokens; seven policy controls preserve the known scope limits. Ruff and format checks pass. Independent review.

[bradhilton]

Minsky review — CLEAR at 9fe17e39db8ad237ae304920b158b69f2f66ebf9 (parent = #885 head da4f165a). Scope is sampled-source identity only; #883's reconcile/manual/repair gaps remain open and no research source adopts this.

Finding: _sampled_evidence_fingerprint previously hashed only output-side evidence (message/text/blocks, token_ids, logprobs, finish/stop reason), so two exchanges with a reused response ID and identical outputs but different captured prompts collapsed into one source key and one causal prefix was dropped. The head adds prompt_token_ids to the fingerprint for all four protocols: Chat and Completions take the choice-level value and fall back to the response-level value when the choice's is missing or null (the same precedence _chat_choice_tokens uses), Responses takes the generation-level prompt_token_ids when present, Messages takes the response extra. Identical captures still hash identically, so legitimate deduplication is unchanged; only distinct causal contexts now stay distinct. The comment on _source_key states the rationale.

Verification at this exact head (ART venv):

• tests/unit/trajectories: 484 passed. Ruff check and format clean on both changed files.
• Negative control: the new test_sampled_source_identity.py against da4f165a → 10 failed / 6 passed. Failing: distinct-prompt binding (all protocols where the response ID is reused), the missing/null choice-prompt fallback cases, and test_reused_response_identity_preserves_both_sampled_causal_prefixes[False/True]. The six that pass on the parent are the identical-capture roundtrip and unrelated-choice controls, which by design do not depend on the change.
• The runtime delta is 15 lines confined to the fingerprint; no tokenizer, renderer, history-splitting or API behaviour changes.

Draft, stacked on #885 and held; McCarthy has cleared this same SHA, so both dedicated-reviewer pass-offs are on 9fe17e39. CI still pending per the thread.