test: prompt_cache_e2e dense multi-turn assertion under-allows for block flooring with thinking models #1156

Problem / Background

prompt_cache_e2e::multi_turn_chat_reports_cached_tokens_and_lowers_prefill_latency (an #[ignore]d test run manually with local qwen3-0.6b-4bit weights) fails deterministically at turn 3: cached_tokens is 64 against a previous prompt length of 73, but the assertion at tests/prompt_cache_e2e.rs:359 allows only 4 tokens of slack.

Attribution is settled by a two-arm measurement: the pre-epic baseline 9c154ff3 fails with token-for-token identical values (cached 0/48/64/96/112 across 5 turns), so this is pre-existing and not a regression from epic #1148. Reproduced 2026-08-15 on main a206f089 and on the baseline with identical numbers; the full turn table is in epic #1148's summary comment.

Mechanism. Dense trie adoption is block-floored to multiples of 16 (all observed cached values are multiples of 16). With max_tokens=16, the qwen3 thinking model produces empty assistant content (all 16 tokens go to the unterminated think block), so the history re-render diverges right at the previous prompt boundary, and flooring can cost up to 15 tokens, which exceeds the 4-token slack.

Proposed Solution

One of:

• Raise the slack to one block minus 1 (15 tokens).
• Derive the slack from the paged block size constant instead of hardcoding 4, so the assertion stays correct if the block size changes.
• Make the fixture avoid the empty-content case: use a non-thinking model, or a larger max_tokens so assistant content is non-empty and the re-rendered history does not diverge at the previous prompt boundary.

The other assertions in the test (cached > 0, monotonic growth) are sound and should stay strict.

Acceptance Criteria

• The multi-turn test passes deterministically with the local qwen3-0.6b-4bit fixture across all 5 turns
• The slack in the under-allowance assertion is either derived from the block size constant or documented with the block-flooring rationale
• The strict assertions (cached_tokens > 0 from turn 2, monotonic growth) remain unchanged

Technical Considerations

Refs epic #1148 (integration verification that surfaced this, full turn table in its summary comment). The deterministic token-identical failure on both arms rules out cache-policy changes from the epic as a cause.