refactor(session-history): move extraction scripts behind skills

[tmchow]

Summary

ce-session-historian used to reach into a sibling session-history-scripts/ directory for its four extraction scripts. That broke the repo convention that scripts live inside skills, not beside agents. This PR moves those scripts into two new agent-internal skills — ce-session-inventory and ce-session-extract — and has the agent invoke them via the Skill tool in fork context.

No behavior change for callers. /ce-sessions and /ce-compound's session-history phase work identically.

Why two skills, not four

The agent uses the scripts in two workflow phases:

Phase	Scripts	New skill
Inventory (always used together, piped)	discover-sessions.sh + extract-metadata.py	ce-session-inventory
Per-session deep-dive (selective, mode-parameterized)	extract-skeleton.py + extract-errors.py	ce-session-extract

One-skill-per-script would have fragmented the primitives below the level the agent operates at. Bundling all four into one skill would have blurred the inventory vs deep-dive distinction the agent makes.

Why user-invocable: false + context: fork

Both new skills set user-invocable: false so they don't auto-fire on user queries. They're agent-internal primitives, not end-user commands.

Both set context: fork so the bash-pipeline plumbing stays in the fork and only the JSONL output bubbles back. The agent's tool-call history stays clean across multiple ce-session-extract invocations during a single research task. The skills pass the "explicit instructions" test fork is designed for: positional args in, stdout verbatim out.

Scope

• Agent rewrite: references to the old script directory removed; extraction delegated via the Skill tool.
• Two new skills under plugins/compound-engineering/skills/, each with a small SKILL.md and a scripts/ directory holding the relevant scripts.
• session-history-scripts/ removed from agents/research/.
• tests/session-history-scripts.test.ts rewired to the new script locations; 30 tests pass.

Smoke-tested end-to-end

Both the /ce-sessions → agent → skills path and direct agent dispatch return clean results on Claude Code:

• ce-session-inventory returned metadata cleanly with parse_errors: 0 on both tests.
• Fork context correctly passed positional args and returned stdout.
• ce-session-extract design verified against the skeleton and errors scripts; not exercised live in the smoke tests (the agent correctly judged metadata alone was sufficient for the two smoke questions).

Conversion notes

user-invocable and context aren't parsed by this plugin's TypeScript skill parser, so they pass through verbatim via copySkillDir when converting to other targets. Claude Code honors both natively. Kiro drops context: fork per its spec. Copilot's writer hardcodes user-invocable: true. End-to-end behavior on non-Claude-Code targets isn't verified yet.

---