refactor(hooks): simplify caching, dispatch flow, and notification helpers

[dgageot]

Same features, less plumbing. Two focused commits split between pkg/hooks and pkg/runtime.

1. refactor(hooks): simplify executor caching and dispatch flow

• Pre-build per-agent hooks.Executor once in NewLocalRuntime (new buildHooksExecutors). The map is read-only afterwards, so the hooksExecMu sync.RWMutex and the double-checked locking dance go away. hooksExec(a) is now a plain map lookup.
• Drop matcher.raw — pattern == nil now means "match all", collapsing matcher.matches() to a single boolean expression.
• Move the "Executing hooks" debug log into Executor.Dispatch so it sits where the work actually happens. Drops the redundant exec.Has(event) early-check in dispatchHook (Dispatch already short-circuits on the empty-event case).
• executeOnUserInputHooks uses r.CurrentAgent() instead of re-implementing the team lookup.
• Remove unused builtins.AgentDefaults.IsZero() (and its test assertion). ApplyAgentDefaults already returns nil on the zero value, which is what callers actually rely on.

2. refactor(hooks): collapse paired notification helpers; extract hooksFor

• executeNotificationHooks + executeOnErrorHooks + executeOnMaxIterationsHooks → notifyError + notifyMaxIterations — two composite helpers that fire the (notification, structured) event pair atomically. The two events are always emitted together in practice, so collapsing them removes 2 lines per callsite in loop.go and surfaces intent ("notify error", "notify max iterations") instead of leaving the reader to mentally join two calls. A private notify() helper shares the (level, message)-shaped Input building.
• Drop the now-dead "invalid notification level" runtime check — the level is always a compile-time constant ("error" / "warning") inside the new helpers, which is stronger enforcement than the runtime branch.
• Extract Executor.hooksFor(event, toolName) from Dispatch, so Dispatch reads top-to-bottom as "collect → serialize → run → aggregate". Drops the redundant len(matchers)==0 early-return; the len(hooks)==0 check now covers both the no-matchers and no-tool-match cases uniformly.

What stays the same

All hook events (12), all hook types (command, builtin), all builtins (add_date, add_environment_info, add_prompt_files), the agent-flag auto-injection, the Claude-compatible Output protocol (both Decision: "block" and HookSpecificOutput.PermissionDecision: "deny"), and the PreToolUse fail-closed security boundary are all preserved.

Validation

• go test ./... — all green
• mise lint — 0 issues
• Reviewed by the PR-review sub-agent: no critical or non-critical issues found, fail-closed security boundary intact, no missed callsites, no race conditions in the new pre-build path.

Diff stat

 pkg/hooks/builtins/builtins.go      |   5 --
 pkg/hooks/builtins/builtins_test.go |   1 -
 pkg/hooks/executor.go               |  58 +++++++--------
 pkg/runtime/hooks.go                | 142 +++++++++++++++---------------------
 pkg/runtime/loop.go                 |   9 +--
 pkg/runtime/runtime.go              |  16 ++--
 6 files changed, 101 insertions(+), 130 deletions(-)

---

Assisted-By: docker-agent