feat: composable context-intelligence behaviors — logging / analytics / full

[colombod]

Intent

The context-intelligence logging hook was doing two unrelated jobs: session-event telemetry AND fetching/refreshing the context-intelligence-graph-query skill content from the server. Those concerns have different consumers — telemetry is wanted by anyone capturing sessions (ideal as an always-on --app), while skill-content sync is only relevant to the analytics path (the graph-analyst agent + graph_query/blob_read tools that actually use the skill). Coupling them forced all-or-nothing composition and put server-fetch + disk-write logic inside a telemetry hook.

This PR separates the concerns: the bundle now exposes three composable behaviors (logging / analytics / full) so consumers pick exactly what they need, and the logging hook becomes pure telemetry.

What changed

1. Skill-content sync relocated: hook → tool-graph-query

• New skill_fetcher.py (server fetch, ETag conditional GET, Bearer auth) and skill_sync.py (on_session_ready orchestration).
• Sync now runs in the graph-analyst sub-session (where the skill is consumed) via the kernel on_session_ready mechanism — not at hook-mount time.
• Offline-drift invalidation: when the server is unreachable and the on-disk SKILL.md hash has drifted, the .etag/.content_hash sidecars are invalidated (content retained) — so a stale ETag can't mask a local reset.
• skill:unloaded re-sync handler for mid-session reloads.

2. Logging hook reduced to pure telemetry

• Removed skill_fetcher.py, legacy_content/, and the skill tests from hook-context-intelligence. Hook __init__.py: 306 → 157 lines.

3. Config-resolver correctness

• Renamed _resolver → _hook_resolver across both tools; the hook resolver is fetched lazily via capability in on_session_ready, not at mount.
• Restored the ToolConfigResolver env/settings fallback for analytics-only mode (its earlier removal would have been a silent breaking change).
• Fixed placeholder expansion: ${AMPLIFIER_CONTEXT_INTELLIGENCE_*} config placeholders are now expanded before the truthiness check, so analytics-only mode resolves the real server URL instead of a truthy literal that silently defeated the env fallback.

4. Dependency hygiene

• All three modules resolve amplifier-core from the PyPI wheel (>=1.6.0) instead of a git/Rust source build — removes a maturin build hang and reverts an out-of-place v1.2.4 downgrade.

5. Documentation

• README: all three behaviors documented with their --app install commands; expanded Development (per-module tests, the venv --reinstall --refresh caveat, DTU pointer).
• Fixed docs/logging-handler-flow.dot rendering and added docs/context-intelligence-skill-sync-flow.dot for the server-sync flow.
• Regenerated bundle.dot/bundle.png for the 3-behavior split.
• Added a contributor AGENTS.md (setup, tests, DTU gate, pitfalls).

Improvements & why they matter

• Granular composition — install logging-only, analytics, or full; pick exactly what you need.
• Separation of concerns — the telemetry hook no longer fetches skills or writes to disk.
• Analytics-only sync actually works (the whole point of the relocation) — proven against a live server.
• No silent staleness — offline drift invalidates sidecars; a reset can't masquerade as a good state.
• Faster, reproducible builds — PyPI wheel, no Rust compile.

Evidence

• Unit tests (fresh): 421 passing, 0 failures — tool-graph-query 87 · tool-blob-read 35 · hook-context-intelligence 299.
• Red-green verified on the core placeholder fix: reverting tool_resolver.py (with uv sync --reinstall --refresh) turns the 3 placeholder tests RED (assert '${AMPLIFIER_…}' == 'secret'); restoring turns them GREEN — the regression test genuinely catches the bug.
• Lint/types clean — ruff (lint + format) pass; pyright 0 errors / 0 warnings on the changed source.
• DTU end-to-end against a LIVE context-intelligence server — all 4 scenarios PASS:
  • S1 analytics-only + reachable server → SKILL.md populated with real server content (MATCH (×44) + .etag/.content_hash written; resolver returned the real URL.
  • S2 offline drift → sidecars invalidated, content retained (validated with frontmatter-preserving drift).
  • S3 logging-only → zero skill activity, telemetry written.
  • S4 full → telemetry + a single skill sync.
• Causality investigation proved the relocated sync's disk writes do not perturb skills discovery; an earlier S2 "failure" was a flawed test (frontmatter-destroying drift), not a product defect.

🤖 Generated with Amplifier

---

Commits and Files Changed

• 21 commits pushed to feat/spike-readonly-behaviour
• 54 files changed: 3395 insertions, 3357 deletions