fix(prompt-cache): stop a whole-prompt hit from duplicating the last token

[inureyes]

Summary

A replay of an identical prompt hits the prompt cache over every prompt token. The adopt restored all len tokens and admission then moved only the prefill cursor back to len - 1, so prefill forwarded the last prompt token onto a cache that already held it. The adopt now stops one token short on both the snapshot and the KV branch, so the re-run lands at its own position.

Hypothesis settled: (a), a real duplicate

Per-layer cache offsets read right after the admission clamp on the pre-fix binary (GB10, CUDA, mlxcel-server, prompt cache on):

Checkpoint	Path	prompt_len	offset after clamp
gemma-3-4b-it-4bit	snapshot, truncating	115	115
gemma-4-12b-it-4bit	snapshot, truncating	119	119
qwen3-4b-4bit	paged KV, prompt padded to 128	128	128
llama-3.2-1b-4bit	paged KV, prompt padded to 160	160	160

The KV path reaches a whole-prompt hit only when the prompt length is a multiple of the paged block (32 by default), which is why #2096 saw it only at those lengths.

What changed

• try_adopt_cached_prefix caps every hit at len - 1 (whole_prompt_reuse_cap). Snapshot branch: the cap turns the restore into a truncating one after asking snapshot_truncatable_to; a refusal (recurrent state, wrapped sliding ring) falls back to cold prefill and records a layout_constraints reject with the entry length. KV branch: the cap runs before the dense truncate and paged block floors, with a prefix_too_short decline if it drops below the minimum.
• The multimodal whole-entry gate keys on the store's match, not the cap, so a whole-entry VLM replay still adopts; its one-token suffix is what the old back-off already forwarded through the token path.
• The admission back-off stays as a warn-level backstop. usage.cached_tokens now reports the restored length instead of one more.

Validation (GB10, CUDA)

Same process, cold request then warm replays, temperature 0, seed 0, 200 tokens:

Checkpoint	Before	After
gemma-3-4b-it-4bit	diverged at char 226	3/3 byte-identical
llama-3.2-1b-4bit	diverged at char 230	3/3 byte-identical
gemma-4-12b-it-4bit	diverged at token 71	diverges at token 27, a bf16 tie (warm top-2 gap 0.0)
qwen3-4b-4bit	diverged at token 76	diverges at token 82, a bf16 tie (warm top-2 gap 0.0)
qwen3.5-0.8b-4bit (recurrent)	identical (boundary hit, not whole-prompt)	identical

The remaining Gemma 4 and Qwen3 divergences are the forward-width effect, hypothesis (b): the warm request forwards 1 (Gemma 4) or 32 (Qwen3, paged block floor) tokens where cold forwards a wider chunk. Control with equal widths (--prefill-chunk-size 1 in both arms): after the fix both are token-identical with bitwise-equal logprobs (max |dlogprob| 0.0); before the fix both still diverge (Gemma 4 differs at the first token, -0.5 vs -0.625).

• Unit: scheduler_whole_prompt_hit_tests (3 tests, tiny Gemma 3) read the restored offset from the model's own snapshot: identical replay, whole-entry hit through admission, and the wrapped-ring refusal falling back to cold. The scheduler prompt-cache suites pass (37 tests).
• e2e: identical_prompt_replay_restores_all_but_the_last_token in tests/prompt_cache_e2e.rs (ignored; qwen3-0.6b-4bit, --apc-block-size 1) asserts cached_tokens == prompt_tokens - 1 and byte-identical replies. It passes here and fails on the pre-fix binary (20 vs 19).
• cargo clippy --release --features cuda --lib --tests -- -D warnings and cargo fmt --check clean.

Not verified on this host

• trinity-nano-preview-4bit (AFMoE) is not present; qwen3-4b and llama-3.2-1b stand in to cover the KV branch, which the issue's checkpoints did not reach.
• The M5 tile-padded broadcast abort from fix(server): pool-backed KVCache::trim is a no-op, so tile-padded prefill leaves pad rows in the cache and default M5 serving returns wrong greedy output #2096 needs Metal hardware; the fix removes its precondition (cache holding len while the mask is built for len - 1).
• A whole-prompt refusal on a real checkpoint was not reachable through chat requests (recurrent families hit their shorter boundary entry instead); the unit test covers that path.

Closes #1760