Fix Gemma 4 session cache across rewritten conversation turns

[TipKnuckle]

Summary

• reconstruct Gemma 4 assistant reasoning from reasoning_content using its native channel markers, so echoed assistant history retokenizes to the committed stream
• retain the pre-decode Gemma 4 prompt boundary and store it with the backend's expected pre-norm hidden state and shared-KV metadata
• restore the longest compatible common prefix when clients rewrite a prior volatile memory suffix, then prefill only the divergent tail
• cover native reasoning token identity and a two-turn memory rewrite with warm/cold state parity

Verification

• python -m pytest tests/test_no_mlx_imports.py tests/test_public_cli.py tests/test_runtime_kpis.py — 268 passed
• focused Gemma/server/session-cache regression set — 375 passed
• python -m build — sdist and wheel built successfully
• scripts/fresh_venv_smoke.sh — passed
• Ruff and git diff --check — passed
• full python -m pytest -q — only five pre-existing exact-float failures in tests/test_graphbank_compiled_verify.py; the same five failures reproduce from clean upstream commit f52f9be

Benchmark Evidence

Measured 2026-08-17 on commit a233e9a:

• hardware: Apple M4 Max, 128 GB, macOS 26.5.2
• model: local Gemma 4 31B QAT target/assistant pair; target mixed 4-bit/8-bit QAT (MLP projections 8-bit), assistant 4-bit
• runtime: resolved sustained profile, MTP generation, reasoning on, SSD session cache off, default Apple fan control
• sampler: temperature 0, seed 1729, max 128 tokens; each compared completion used 53 tokens
• workload: second OpenAI-format turn removes the prior user's volatile <memory_context>, echoes assistant reasoning_content plus content, and injects a new memory block on the current user turn

Run	Prompt	Cached	New prefill	Prompt eval	Wall
Warm rewritten turn	1,705	1,610	95	0.519 s	2.028 s
Cold control	1,705	0	1,705	8.436 s	9.944 s
Warm exact repeat	1,705	1,705	0	0.0001 s	1.545 s

The rewritten warm turn was 4.9x faster end-to-end than cold; the exact repeat was 6.44x faster. The complete returned message, including both reasoning_content and visible content, matched the cold control exactly in both warm cases.