feat(mcp): codegraph_explore as the sole primary tool + store coverage + overload disambiguation (#647)

## Summary

Completes the explore-overhaul arc: `codegraph_explore` becomes the single primary tool an agent reaches for, and its coverage + output shape are tuned so flow/architecture questions resolve with near-zero Read/Grep.

### What changed
- **explore is the sole primary tool** — removed `codegraph_context` (the fuzzy-input Read-trigger) and `codegraph_trace` (under-picked by agents); explore already surfaces the call flow among the symbols you name. A plain natural-language question now works as the query.
- **Store/handler coverage** — functions defined inside object literals (Zustand `create((set, get) => ({ … }))`, Redux/Pinia/MobX, exported handler/route maps) are indexed as real symbols, including calls through `useStore.getState().fn()` and destructured `const { fn } = useStore.getState()`. A general AST rule, not a per-lib hack.
- **Overload disambiguation** — explore leads with the *right* definition when a method name is overloaded across types (a PascalCase type token in the query biases to that type's own def); `codegraph_node` returns *every* overload's body in one call, with an optional `file`/`line` selector to pin one.
- **Method-atomic render** — explore never returns half a method; at the size budget it drops whole methods/files (and lists what it dropped) instead of truncating a body mid-method.
- **Native-read-shaped output** — per-call output is capped to ~24K with a 25K hard ceiling and concentrated into ~150–250-line flow windows, mirroring how the agent natively reads; repo size scales the *call* budget, not the per-call size (a larger response just gets externalized to a file the host Reads back).
- **Blast radius** folded into explore (dependents + covering tests, locations only).

### Benchmark (refreshed on this build)
Re-validated the 7-repo A/B on 2026-06-02 (Opus 4.8, effort=high, median of 4). WITH arm re-measured on this build, WITHOUT reused:

**~16% cheaper · 47% fewer tokens · 22% faster · 58% fewer tool calls** — 0 file reads on 6 of 7 repos (Gin ~1).

The arc trades larger, cache-heavy explore responses for guaranteed near-zero reads, so cost/token margins soften vs the prior build (Excalidraw and Tokio land at cost break-even) while time and tool-calls stay clear wins everywhere — consistent with the project's stated optimization target (latency + tool-calls, not token cost).

### Validation
- Full suite green: **1112 passed, 2 skipped**.
- 28/28 plain WITH runs across the 7 README repos completed clean; reads median 0 on 6/7.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

CLAUDE.md

@@ -102,12 +102,12 @@ The lever that decides whether a retrieval change lands. **Test before building
 CodeGraph's only channels to influence the agent are low-salience: the MCP `initialize` instructions (`server-instructions.ts`) and the tool descriptions. Changing them does **not** reliably move the agent's tool _choice_ or query style — validated: trace-first steering ported into the server-instructions + tool descriptions (3 wording variants) never reproduced what a CLI `--append-system-prompt` achieved, and **regressed** wall-clock vs baseline. New tools fare worse (rarely chosen — the agent under-picks even `trace`); "better examples" is the same steering. The agent's tool-choice does improve on its own as host models get better at tool use — but that is not ours to force.
 
 What works is meeting the agent where it already is:
-- **Sufficiency** — `codegraph_trace` inlines each hop's body + the destination's own callees, so one trace call ends the flow investigation (no follow-up explore/node/Read).
-- **explore-flow** — `codegraph_explore`'s query is a precise bag of symbol names (incl. qualified `Class.method`) spanning the flow the agent is after; explore finds the call path _among those named symbols_ (riding synthesized edges) and leads its output with it — delivering trace-quality flow through the call the agent reliably makes. (`buildFlowFromNamedSymbols`: segment/co-naming disambiguation; ≤1 unnamed bridge so it never wanders a god-function's fan-out.)
+- **explore-flow** — `codegraph_explore` is the PRIMARY tool the agent reliably calls; its query is a precise bag of symbol names (incl. qualified `Class.method`) spanning the flow the agent is after; explore finds the call path _among those named symbols_ (riding synthesized edges) and leads its output with it. (`buildFlowFromNamedSymbols`: segment/co-naming disambiguation; ≤1 unnamed bridge so it never wanders a god-function's fan-out. Overload-aware: a PascalCase type token in the query biases an overloaded name to that type's own def — `DataRequest task` → DataRequest's `task`, not the abstract base; named-symbol files sort first.)
+- **Sufficiency** — make the tool's output complete enough that the agent stops. `codegraph_node` returns the full body + the caller/callee trail, and for an AMBIGUOUS name returns **every overload's body in one call** (so the agent never Reads a file to find the right overload — validated on Alamofire/gin). This is the after-explore depth tool (labeled SECONDARY).
 
-What fails is the inverse — folding a precise answer into a **fuzzy-input** tool. `codegraph_context` gets a description, not symbols, so it can't disambiguate a flow's endpoints and surfaces the _wrong feature_. Precise output needs precise input.
+What fails is the inverse — folding a precise answer into a **fuzzy-input** tool: the now-removed `codegraph_context` took a description, not symbols, so it couldn't disambiguate a flow's endpoints and surfaced the _wrong feature_ (which is why it was cut). Precise output needs precise input — explore takes a symbol bag for exactly this reason. (`codegraph_trace` was likewise removed: explore-flow does its job and the agent under-picked it.)
 
-The remaining lever under this axis is **coverage**: every flow made to connect statically (a new dynamic-dispatch synthesizer) is then surfaced automatically by explore-flow/`trace`, no agent change needed. Reactive/reconciler runtimes (Halo's `ReactiveExtensionClient`, MediatR, Vue Proxy) are the frontier — flows there have no static edges, so nothing surfaces (correctly — silent beats wrong). Full investigation + A/B record: `docs/benchmarks/call-sequence-analysis.md`.
+The remaining lever under this axis is **coverage**: every flow made to connect statically (a new dynamic-dispatch synthesizer, or extracting symbols static parsing skipped — e.g. object-literal store actions in `create((set,get)=>({...}))`) is then surfaced automatically by explore-flow, no agent change needed. Reactive/reconciler runtimes (Halo's `ReactiveExtensionClient`, MediatR, Vue Proxy) are the frontier — flows there have no static edges, so nothing surfaces (correctly — silent beats wrong). Full investigation + A/B record: `docs/benchmarks/call-sequence-analysis.md` + auto-memory `project_codegraph_read_displacement`.
 
 ### Explore budget — keep BOTH budgets monotonic with repo size
 
@@ -126,7 +126,7 @@ Two functions in `src/mcp/tools.ts` scale explore with indexed file count. This
 
 ### Dynamic-dispatch coverage — the flow must EXIST in the graph end-to-end
 
-Static tree-sitter extraction misses computed/indirect calls, so flows break at dynamic dispatch and the agent reads to reconstruct them. Synthesizers/resolvers bridge these so `trace`/`explore` connect end-to-end (`src/resolution/callback-synthesizer.ts`, `src/resolution/frameworks/`). Channels today: callback/observer, EventEmitter, **React re-render** (`setState`→`render`), **JSX child** (`render`→child component), django ORM descriptor. All synthesized edges are `provenance:'heuristic'` with `metadata.synthesizedBy` + `registeredAt` (the wiring site), surfaced inline in `trace`, the `node` trail, and `context` call-paths.
+Static tree-sitter extraction misses computed/indirect calls, so flows break at dynamic dispatch and the agent reads to reconstruct them. Synthesizers/resolvers bridge these so `codegraph_explore` connects them end-to-end (`src/resolution/callback-synthesizer.ts`, `src/resolution/frameworks/`). Channels today: callback/observer, EventEmitter, **React re-render** (`setState`→`render`), **JSX child** (`render`→child component), django ORM descriptor. All synthesized edges are `provenance:'heuristic'` with `metadata.synthesizedBy` + `registeredAt` (the wiring site), surfaced inline in `codegraph_explore`'s Flow section and the `codegraph_node` trail.
 
 **Principle: partial coverage is WORSE than none.** Bridging one boundary but not the next reveals a hop the agent then drills + reads to finish. Measured on excalidraw: react-render alone *raised* reads to 5–7; only completing the flow (adding the jsx-child hop) dropped it to 0–1. **Always close the flow end-to-end and re-measure** — never ship a half-bridged flow.
 
@@ -135,7 +135,7 @@ Static tree-sitter extraction misses computed/indirect calls, so flows break at
 For each **language × framework**, validate on **small, medium, and large** real repos with **≥3 different flow prompts** each:
 
 1. **Pick the canonical flow** for the framework ("how does X reach Y": state→render, request→handler→view, query→SQL, action→reducer→store…).
-2. **Deterministic probes** (`scripts/agent-eval/probe-{trace,node,context,explore}.mjs` against the built `dist/`): `trace(from,to)` connects end-to-end with no break; **no node explosion** (`select count(*) from nodes` stable before/after re-index); synthesized-edge **precision** spot-check (`select … where provenance='heuristic'`).
+2. **Deterministic probes** (`scripts/agent-eval/probe-{node,explore}.mjs` against the built `dist/`): `codegraph_explore` with the flow's symbol names connects from→to end-to-end with no break (its Flow section shows the path); **no node explosion** (`select count(*) from nodes` stable before/after re-index); synthesized-edge **precision** spot-check (`select … where provenance='heuristic'`).
 3. **Agent A/B** (`scripts/agent-eval/run-all.sh <repo> "<Q>"`): with vs without codegraph, **≥2 runs/arm** (run-to-run variance is large — never conclude from n=1). Record **duration, total tool calls, Read, Grep**. Optional forced-Read-0 sufficiency proof via the block-read hook (`scripts/agent-eval/hook-settings.json`).
 4. **Pass bar:** a normal flow question reaches **~0 Read/Grep within the repo's explore-call budget**, runs **faster** than without-codegraph, and shows **no regression on a control repo**. Record the numbers in `docs/design/dynamic-dispatch-coverage-playbook.md` (the coverage matrix).
 
@@ -209,7 +209,7 @@ Formatting rules for any entry (anywhere — `[Unreleased]` or otherwise):
 
 1. **Write friendly, user-facing notes — not engineer-facing ones.** Group under `### New Features` and `### Fixes` (sentence-case). Surface `### Breaking Changes` and `### Security` as their own sections **only when the release has them**; fold improvement-flavored changes into New Features. Omit empty sections. (This replaces the old Keep-a-Changelog `Added/Changed/Fixed/Removed/Deprecated` grouping: the GitHub Release page extracts each version block **verbatim** via `scripts/extract-release-notes.mjs`, and the old dense, implementation-focused entries rendered as an unreadable wall of text — so the whole CHANGELOG was rewritten to this format and every published release re-noted to match.)
 2. **One plain-language sentence per bullet:** what changed and why it matters to a user. Lead with the capability, or with the symptom that's now fixed.
-3. **Strip the internals.** No internal file paths (`src/...`), no internal symbol / function / class names, no benchmark numbers / percentages / node-or-edge counts. **Keep:** language & framework names (Go, Spring, NestJS, …), things a user types or sets (`codegraph install`, `codegraph_trace`, the `CODEGRAPH_*` env vars), agent / IDE names (Claude Code, Cursor, opencode, Kiro, …), and a brief `Thanks @user` when a contributor is credited.
+3. **Strip the internals.** No internal file paths (`src/...`), no internal symbol / function / class names, no benchmark numbers / percentages / node-or-edge counts. **Keep:** language & framework names (Go, Spring, NestJS, …), things a user types or sets (`codegraph install`, `codegraph_explore`, the `CODEGRAPH_*` env vars), agent / IDE names (Claude Code, Cursor, opencode, Kiro, …), and a brief `Thanks @user` when a contributor is credited.
 4. Issue / PR references in entries are by number (`(#403)` etc.); the GitHub renderer auto-links them in the published release notes.
 5. **Don't add a `[X.Y.Z]: https://...` link reference yourself** — `prepare-release.mjs` appends it automatically when it promotes the version (idempotent: a re-run is a no-op if it already exists).