fix(claude-code): replace hardcoded hook with rich episodic logging

[aliirz]

The PostToolUse hook was calling:
memory_reflect.py claude-code post-tool ok
for every tool call, producing identical entries with action='post-tool', detail='ok', and reflection=''. content_cluster() has nothing to cluster on, so the dream cycle staged zero candidates regardless of how many sessions ran.

Changes

.agent/harness/hooks/claude_code_post_tool.py [NEW]
Reads the JSON payload Claude Code sends via stdin on every PostToolUse
event (tool_name, tool_input, tool_response). Falls back to
CLAUDE_TOOL_NAME / CLAUDE_TOOL_INPUT env vars for older versions.

• Derives a meaningful action label from tool name + input
• Scores importance from OPERATION (deploy/migration/schema = 9, commit/test/build = 6, routine = 3) not from service brand names
• Detects failures via exit_code, error stream, interrupted, is_error
• Generates a non-empty reflection the dream cycle can cluster on
• Sets pain_score=5 for high-importance successes so recurring patterns cross the 7.0 promotion threshold (was always 2, max salience = 5.4, never reached 7.0)

.agent/protocols/hook_patterns.json [NEW]
User-owned config for stack-specific service names. Ships with empty
high_stakes / medium_stakes arrays and a populated _examples section.
Users copy from _examples to activate (e.g. 'supabase', 'vercel').
The hook merges these with the universal built-in patterns at startup.
Universal patterns (deploy, migration, schema, production, secret, etc.)
stay hardcoded; service names do not.

.agent/harness/hooks/post_execution.py
Add optional pain_score parameter to log_execution(). Default behaviour
unchanged (2 for success, 7 for failure).

.agent/harness/hooks/on_failure.py
When error is a plain string, omit the noisy 'str:' type prefix from
the reflection field. Only add type name for actual Exception objects.

.agent/tools/memory_reflect.py
Expose pain_score via --pain CLI flag so manual reflect calls can set
it explicitly (e.g. --pain 5 for a significant successful discovery,
--pain 9 for a production incident).

adapters/claude-code/settings.json
Point PostToolUse hook at claude_code_post_tool.py. Expand matcher to
include MultiEdit, Task, TodoWrite. Add supabase db reset to deny list.

adapters/claude-code/CLAUDE.md
Add explicit 'when to log manually' section with examples, importance
guide table, and good/bad call examples.

adapters/claude-code/README.md
Document the new hook, why the old one was broken, how to verify,
and common troubleshooting scenarios.

test_claude_code_hook.py [NEW]
54-assertion validation suite. Run from any project that has .agent/
installed. Exit 0 = all good. Tests cover: import, empty stdin,
action labels, importance scoring, pain calibration, failure detection,
reflection non-empty, full write path, dream cycle integration,
--pain CLI flag, hook_patterns.json loading and overrides.

Fixes

• dream cycle always produced staged=0 (empty reflections)
• successful high-stakes ops never auto-promoted (salience ceiling 5.4 < 7.0)
• failures silently logged as success (no exit code / error stream check)
• all tool calls uniform importance=3 (no domain awareness)
• service names hardcoded in importance scorer (not portable across stacks)