v1.108.149 — deterministic emission + cache-stability baseline

Two things: an ordering guarantee, and a measurement that honestly declined to justify a bigger feature.

Deterministic emission. Six context-serving sorts (get_ranked_context exact-seed and scoring paths, get_repo_map file ranking, signal-fusion, assemble_task_context anchor and plate picks) resolved tied float scores by list insertion order — which traces back to index storage order and can shuffle across rebuilds of an unchanged tree. Ties now break on symbol id / file path, so result ordering is a function of content and scores alone: two indexes built from the same tree in different discovery orders serve identical rankings. Positions driven by real score differences are unchanged (Replay-gated, green).

Cache-stability baseline. Provider prompt caches bill cached reads at ~0.1x but key on byte-identical prefixes — so re-shuffled repeat context is never cache-eligible. The new benchmarks/cache_stability/ harness measures how much of what get_ranked_context serves across realistic drill-down sequences is repeated symbols in a different order. Measured against a pinned self-index: 15.8% of served bytes are repeats, 13.9% re-shuffled — below the pre-registered 30% go/no-go threshold for building a byte-stable emission mode. So that feature is deliberately on hold, with the number committed in results.json. Pre-registering the threshold and publishing the miss is the point: measurements you only report when they justify the feature aren't measurements.

Full suite: 5413 passed.