feat: ADR-018 AC-3 cross-turn prefix reuse (v0.3.11) + full-suite benchmark

[Tovli]

Summary

Performance phase 0.5: lands ADR-018 AC-3 (cross-turn prefix reuse / incremental prefill), bumps the workspace to 0.3.11, and documents the measured win with a full-suite el-bench report plus a README progression table.

Three focused commits:

Commit	What
feat(runtime,engine): ADR-018 AC-3	Reuse the KV cached for the longest matching token prefix and prefill only the divergent suffix, instead of re-prefilling the whole conversation each turn.
chore: bump workspace version to 0.3.11	cargo set-version --workspace 0.3.11 (follows master's 0.3.10).
docs(benchmarks): AC-3 report + README	New 2026-06-23 report and a benchmark progression table in the README.

What AC-3 changes

• el-runtime — InferenceSession::continue_prompt + the InferenceEngine::prefill_reuse port. Safe default re-prefills the full context; stateful engines override for the fast path. A failed continue leaves a recoverable session.
• el-engine-candle — QwenEngine::prefill_reuse does longest-common-prefix reuse against the live cache and falls back to a full prefill on any mismatch.
• Soundness — reuse must leave the cache byte-identical to a fresh prefill: proven by the ReuseEngine mock test prefill_reuse_leaves_same_state_as_a_fresh_prefill, and it fails open on a wrong/short match.

Measured impact (full el-bench suite, Qwen2.5-0.5B q4_k_m, i5-14500)

Same harness/inputs/host as the 2026-06-22 run; only the AC-3 code differs. 66 tasks → 97/97 replies, 0 errors.

Metric	2026-06-22 (no AC-3)	2026-06-23 (AC-3)	Change
Full-suite wall	3221 s (53.7 min)	2141.5 s (35.7 min)	−33.5%
Prefill share of compute	56.8%	29.5%	−65% wall
Prefill forwards (27,148 prompt tok)	~27,148 (1/token)	9,893	64% served from cache
mindeval (multi-turn) avg ms/reply	52,225	17,478	~3× faster

Honest caveat (in the report): the prefill kernel is unchanged — per-forward rate is still ~15 tok/s — so AC-3 wins by doing less work, not faster math. ADR-020 (batched prefill) is still pending for cold/first-turn prefill, and the ~14 tok/s decode floor is now the dominant cost.

Verification

• cargo build --workspace — clean at 0.3.11 (excluding el-chat, whose binary was locked by a stale local process; el-chat source is untouched by this PR and compiled fine).
• cargo test -p el-runtime — 31/31 pass, including the AC-3 soundness and recoverable-session tests.
• The benchmark itself ran end-to-end through the real SDK path with 0 errors and matching completion-token counts vs the prior run (AC-3 changed speed, not output).

Notes for reviewers

• The AC-3 implementation (session.rs, ports.rs, el-engine-candle/src/lib.rs, ADR-018 doc) was already present in the working tree at the start of this work; this PR commits it as-is alongside the version bump and benchmark docs.
• The branch was rebased onto current master (which had advanced to 0.3.10) before committing, so the diff is clean — no spurious revert of master's README/CI/packaging commits.
• The prior 2026-06-22 report's "prefill is the Add multi-registry release CI (crates.io · npm · pub.dev) — ADR-011 #1 cost" conclusion is now superseded on the prefill axis by the new report (left intact for history).

🤖 Generated with Claude Code