feat(kimi): full Kimi Code CLI hook support (PreToolUse deny, Stop, subagent briefing)

[zzet]

Closes #214.

Kimi Code CLI support was previously UserPromptSubmit + a read-tool nudge only. This brings it to parity with the Claude Code and Codex adapters, so weaker-instruction-following models get the full deterministic hook loop the issue asks for: context injection, tool redirection, and post-turn correction.

What's added

Event	Behavior
UserPromptSubmit	(existing) injects graph symbols relevant to the prompt before the model runs.
PreToolUse	new — native Read/Grep/Glob/Bash run through the shared enrich machinery: a hard deny for an indexed whole-file read, soft plain-stdout guidance otherwise; keeps the read_file/get_editing_context compress-bodies nudge.
Stop	new — post-turn diagnostics (changed symbols → test targets, guards, dead code, coverage, contracts), fed back so the agent self-corrects before handoff.
SubagentStart	new — briefs a spawned subagent with smart_context results and the tool-swap table so it doesn't default to raw Read/Grep.

Wire protocol

Kimi cloned Claude Code's hook schema, so a hard block uses the one JSON shape Kimi documents — hookSpecificOutput.permissionDecision = "deny" — and all soft guidance rides Kimi's plain-stdout context channel (matching the established Kimi contract). Every path stays gated to Gortex-enabled projects and degrades to a silent no-op when the payload is malformed, the cwd is outside a project, or the daemon is unreachable — so normal Kimi flow is never blocked by the integration.

Transport fix (why the Stop path is non-vacuous)

callServerTool (used by the Stop and subagent diagnostics) only spoke the HTTP :8765 surface, which the daemon serves only under --http-addr — so it was dead by default. It now falls back to the daemon's AF_UNIX socket, scoped to the hook cwd. The fallback is gated on a recorded cwd, so the pure-HTTP unit tests are unaffected and non-Kimi agents keep their current behavior.

Not included

SubagentStop is intentionally left out — the Stop hook already covers end-of-turn diagnostics, and a per-subagent variant would mostly add noise. Easy to add later if it proves useful.

Verification

• go build ./..., go vet, gofmt, golangci-lint — all clean.
• go test -race ./internal/hooks/... ./internal/agents/... — 527 pass (new tests cover the native-tool deny JSON, soft-stdout guidance, Stop diagnostics, subagent briefing + fallback, project gating, and the transport fallback gating).
• go test ./cmd/gortex/ -run 'AgentRender|Hook|Kimi|Install|Init' — 29 pass.
• Docs: docs/agents.md adapter matrix + per-agent ### kimi note added.