feat(skill): add context-intelligence-hill-climbing skill for audit-trail tracking

[colombod]

Summary

Add a thin, shared context-intelligence-hill-climbing skill (todo-as-climb-log) and wire both graph-analyst and session-navigator agents to load it on multi-step investigations. An earlier evaluation found both agents already navigate safely and correctly (bounded probes, no context flooding) but leave an inconsistent audit trail — baseline session-navigator used 0 todo calls while still answering. This skill's purpose is observability/consistency, NOT correctness or efficiency.

Scope / guardrails

• Modified: agents/graph-analyst.md (added "Track multi-step investigations in a todo climb-log" guidance block)
• Modified: agents/session-navigator.md (identical wiring block)
• New: skills/context-intelligence-hill-climbing/SKILL.md (thin skill, ~2 KB, focused on using todo as climb-log)
• NOT touched: write-side hook (logging_handler.py, _DestinationDispatcher, fanout.py), navigation discipline mechanics (those stay in graph-query, session-navigation, navigation-budget-discipline.md), bundle structure

The skill governs HOW an investigation is tracked (seed open questions, write findings + citations, mark dead leads) — it deliberately does NOT re-teach navigation mechanics.

Verification

• Module tests pass — modules/tool-context-intelligence-query: 132 passed
• Top-level tests pass — tests/: 18 passed
• ruff check + ruff format --check clean — 0 errors
• pyright clean — 0 errors
• Full bundle validation PASS — scripts/validate-full.sh → validation_mode: full
  The lone mode-advertising ERROR is a documented FALSE POSITIVE (name collision between bundle name, storage path, skill name, and internal mode) — see AGENTS.md. Validator confirmed overall verdict PASS.

Real evidence on seams (not mock-only)

This change crosses a seam: agent wiring (skill loading). Per AGENTS.md, seam changes require real DTU run + evaluation harness with captured evidence.

Lightweight A/B evidence (delegate-based, n=1 per cell, skill injected as inline guidance because not yet installed):

Both agents were asked the same real RCA/provenance question (session 1ade4f66…, which session produced the auto-recovery design first?) on both surfaces (graph-analyst, session-navigator), cold baseline vs. +skill treatment.

run	total tool calls	todo calls	correct?	audit trail
graph-analyst baseline	10	6	yes	ad-hoc
graph-analyst +skill	18	10	yes	structured
session-navigator baseline	31	0	yes	none
session-navigator +skill	46	8	yes	structured

The +skill runs produced a structured, cited todo climb-log (open questions seeded up front; findings written into items with node-id/line citations). Correctness held at baseline and remained correct with the skill. Cost rose ~50–80% (expected overhead for an observability feature).

Session-navigator's final climb-log from the +skill run read:

✓ Resolve root session dir and confirm sub-session locations
✓ Q1 ROOT CAUSE — ANSWERED: foundation:explorer sub-session (…7bc831ed…, spawned 15:48:58Z from root events.jsonl:162)…
✓ Q2 FIRST DESIGN AUTHOR — ANSWERED (nuanced): design authored directly by the ROOT session itself, not a delegated sub-agent…
✓ Q3 CAUSAL CHAIN — ANSWERED: error report → foundation:explorer → root-session brainstorming+ROB → council → brainstormer (documents)…

• Seam(s) crossed by this change are proven against real behavior (lightweight A/B on real scenario; see above)

NOTE: This repo's AGENTS.md requires agent/skill/mode edits to pass "a real DTU run + the evaluation harness, with captured evidence" — not just lightweight delegate tests. That full DTU-isolated, multi-trial run has NOT been executed. RECOMMEND running the DTU + evaluation-harness gate before merge (see .amplifier/digital-twin-universe/profiles/ for available profiles).

Docs & diagrams

• bundle.dot / bundle.png regenerated if bundle structure changed — N/A — no bundle structure change (skill addition is additive, no new seams in bundle.md or behaviors layout)
• README / SKILLs / agent files updated if a tool/skill contract changed — documentation in skill SKILL.md and agent guidance blocks reflects the new capability
• Convention files updated if this surfaced a lasting lesson — N/A — no new convention learned

Notes / follow-ups

Gate not fully met: This PR provides lightweight evidence (delegate-based, n=1 per cell) that the skill produces structured audit trails without breaking correctness. Full DTU + evaluation harness validation remains as a pre-merge recommendation per AGENTS.md § Testing & what "done" looks like. The DTU profiles exist in .amplifier/digital-twin-universe/profiles/ (e.g., context-intelligence-bundle-smoke-test.yaml) — running one of them with this branch's agents will give the full proof required before shipping.