Add the disk-hygiene plugin for session-owned cleanup

[cdeust]

Summary

• Adds plugins/disk-hygiene, a shared cleanup hook for Claude Code and Codex, moved out of fix(hooks): share cleanup across Claude and Codex, retain transcripts Cortex#645 because cleanup of session files does not belong to a memory server.
• The removal logic is unchanged. Fixed from the review of that PR: push detection reads only the command of a shell tool; the ledger lock waits up to 10 seconds and is not taken by ordinary tool calls; DISK_HYGIENE_CLEANUP=off disables every hook; the context-guard checkpoint follows the transcript policy; a corrupt or foreign end-of-session intake file is skipped and reported; a symlinked config directory is resolved; the Cortex completion receipt for Codex rollouts is optional.
• Transcripts are kept by default. Deleting them is opt-in.
• Refs MCP server holds a multi-GB memory plateau per session; orphaned servers survive their client #47. This does not fix the memory plateau measured there, which still needs a profile of one server.

Test plan

• 132 tests for the new plugin pass locally, including false-positive push detection, lock contention, the global switch and the invalid intake case.
• CI on this branch.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Mkvy5MtGG89cNLuQSU8xLe

[cdeust]

ZETETIC-REVIEW: REQUEST_CHANGES

Independent read-only review of 3d6ca63 and 42977e3 against the original files of cdeust/Cortex#645 (head 9a8b606b). No wrongful deletion path was found: cleanup_operations, session_purge and codex_purge differ only on the seven intended points plus source wording, import order and an explicit check=False; the guards in remove() (symlink, owner, nested .git, lsof, transcript reader) are intact; session ids pass the UUID check before any path or glob is built.

Findings:

1. Medium. An intake with a valid shape but a non-UUID session id was written at SessionEnd, then made the host import raise on every later hook, blocking all host cleanup. Not executed by the reviewer, reasoned from the code.
2. Low. The README said a foreign intake is reported on stderr; it is filtered silently and left in place.
3. Low. The ledger location and the CORTEX_CLAUDE_DIR variable changed without a migration note.
4. Low. status fails on an invalid DISK_HYGIENE_CLEANUP value; consistent with the README, left as is.

Not verified by the reviewer: no test was run, and the real Codex tool_name values were not compared with SHELL_TOOLS.

[cdeust]

ZETETIC-REVIEW: APPROVE

Follow-up to the REQUEST_CHANGES verdict above, on head 4562db8. Findings 1 to 3 are fixed and finding 4 is accepted as documented.

• Finding 1 (non-UUID session in the host intake): read_pending now reports and skips such a file for the host consumer and still returns it for the worktree consumer. test_non_uuid_session_intake_does_not_block_host_cleanup covers both consumers. The existing lifecycle test that relied on the old behavior now asserts that the file stays on disk.
• Findings 2 and 3: the README states what happens to an intake recorded under other roots, and the README and CHANGELOG document the ledger move and the CORTEX_CLAUDE_DIR to CLAUDE_CONFIG_DIR rename.
• Evidence: 253 tests pass locally, ruff 0.16.8 check and format are clean, and CI is green on 4562db8 (test, CodeQL, Analyze).
• Still not verified: the real Codex tool_name values against SHELL_TOOLS, which needs a native Codex run after installation.