feat: surface code-graph tools so the agent stops forgetting them (3-layer)

[pirumu]

Make the agent actually reach for the code-graph tools

The graph tools (query/paths/communities/processes/taint) work over MCP + CLI, but nothing led the agent to them — it kept grepping. Documentation alone gets forgotten. This wires them in structurally, at three layers of increasing forget-resistance, matching the repo's proven anti-forgetting pattern (hooks that fire at the moment of action, like lesson-guard).

Layer 1 — sharpen MCP tool descriptions (highest leverage, zero-recall)

The tool list is the one surface the agent never forgets — it's re-presented every turn. The graph tools described their mechanism ("BFS traversal from FQNs", "label propagation") but not the situation that should trigger them. Rewrote all five to lead with the use-case and, where it applies, "reach for this INSTEAD of repeated Grep": trace-how-code-connects (query), A→B path (paths), architecture map (communities), execution flow (processes), security dataflow (taint).

Layer 2 — behavior-triggered graph-affordance PreToolUse hook (catches the agent mid-forget)

A non-blocking PreToolUse hook on Grep|Glob that counts code-searches in the session's enforcement log and, once past a threshold (4), surfaces a one-time nudge listing the graph tools. Fires once per session (sentinel, wiped at SessionStart so it re-arms), never blocks — a nudge, not a gate (the graph tools are optional-better, not mandatory, so blocking would be wrong). Same architecture as lesson-guard / the frustration prompt-guard. Pure counting/decision helpers are unit-tested; hook-registration count assertions updated.

Layer 3 — docs (fallback layer)

Added the graph tools to memory-guidance (the recall-before-acting list), the memory-exploring skill (mapped to the questions that already live there — architecture → communities, flow → processes, A→B → paths), and the memory-cli skill (new graph command family). This is the weakest layer on its own, included so it helps when the agent does read a skill.

Verification

• Hook smoke (temp project): silent on grep 1–3, nudge on grep Chore/rename thoth to hoangsa memory #4, silent on fix(memory): 4 memory-quality bugs — FQN collision, :: query, init lock, chroma default-on #5 (once-per-session), silent again after state-clear (re-arms).
• cargo test on touched crates 0 failed, clippy 0 warnings, attribution gate 0.

Why not a hard block

require-memory-impact blocks because impact-before-edit is mandatory. Using the graph is a preference, not a requirement — so nudge, never block.