feat(prompt-hook): graph-derived gate tier + confidence-tiered injection + gate telemetry (#1136)

The keyword gate (#1126) can never know a repo's domain nouns. This adds
the graph-derived tier the design discussion converged on: symbol names
are split into prose segments at index time (name_segment_vocab, riding
the insertNode write path), and the hook verifies a prompt's plain words
against them — "the state machine des commandes" → OrderStateMachine, in
any language whose technical nouns are Latin script.

Confidence now decides HOW MUCH to inject, not just whether:
- HIGH (keyword, or index-verified code token): full explore injection,
  unchanged — the validated adoption lever.
- MEDIUM (segment matches only): a ~500-byte pointer naming the matching
  symbols; the AGENT writes the explore query. Never runs explore, so a
  fuzzy match can't inject 16KB of wrong-feature context.
- Silent otherwise, as before.

Precision is derived from the repo's own naming statistics plus measured
FP fixes: co-occurrence (≥2 words on one name) always qualifies; a single
word must be ≥5 chars, cluster across 2–25 names (singletons are prose
coincidence: "deploy to production" → matchesNonProductionDir), match a
multi-segment name, and not be an English function/filler word (the one
place a word list is honest: identifiers are English, so only English
prose collides). Every candidate is re-verified against nodes before
being surfaced — vocab rows are proposals, deletions leave orphans by
design, a full index rebuilds from scratch, and sync heals pre-upgrade
databases (batched + yielding; emptiness captured at sync ENTRY so the
sync's own writes can't mask the backfill).

Schema v7 migration is DDL-only (instant; none of the #1067 row-churn
hazards). Gate outcomes roll up as anonymous usage counters
(prompt-hook-gate-<outcome>, names only, never content) through the
existing telemetry pipeline — recall becomes measurable, and the counters
are the agreed kill-criterion data for ever revisiting a local classifier.

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

CHANGELOG.md

@@ -9,6 +9,11 @@ and adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 ## [Unreleased]
 
+### New Features
+
+- The Claude Code context hook now recognizes prompts that describe code in plain words — in any language — by checking the prompt's words against the symbol names actually in your project's index. Asking about "the state machine des commandes" finds `OrderStateMachine` with no keyword involved. Confidence decides how much gets injected: structural questions and prompts naming a real symbol still get full context up front; a plain-words match gets a short pointer to the matching symbols so the agent queries them itself; everything else stays silent, exactly as before.
+- Anonymous usage telemetry now counts how often the context hook injected context, offered a hint, or stayed silent — fixed counter names only; the prompt's content is never stored or sent. This makes the hook's accuracy measurable instead of guessed.
+
 ### Fixes
 
 - The automatic context hook for Claude Code now fires for structural questions asked in nearly thirty languages — French, Spanish, Portuguese, German, Italian, Dutch, Polish, Czech, Romanian, Hungarian, Greek, Swedish, Danish, Norwegian, Finnish, Russian, Ukrainian, Turkish, Indonesian, Vietnamese, Thai, Hindi, Arabic, Farsi, Hebrew, Japanese, Korean, and both simplified and traditional Chinese — instead of just English and simplified Chinese. Previously a natural question like "comment marche la state machine des commandes ?" injected nothing unless it happened to contain a code-shaped symbol name, making the hook look broken for non-English teams. English questions phrased with derived word forms ("explain the architecture…", "what are the dependencies…") now fire too, and prompts in any other language still fire when they name a symbol from the index. Thanks @anthonyle-roy-lgtm for the report. (#1126)