Migration notes for akm v0.8.0

This release combines the 0.8.0 CLI/storage break with the final improve-owned maintenance migration work that landed before release validation closed.

Key operator-facing changes:

• akm index --enrich and akm index --re-enrich are removed.
• Plain akm index now owns metadata enhancement only.
• Memory inference and graph extraction now run from akm improve after consolidation, not from akm index.
• Graph schema redesign (DB_VERSION 12 → 13): graph_files re-keyed on entry_id with FK cascade to entries(id), child tables follow. Upgrade drops the graph tables and repopulates them on the next akm improve cycle; non-graph tables rebuild automatically. Run akm improve once after the storage migration to repopulate graph data.
• Manual QA and sandbox guidance now require isolating AKM_DATA_DIR in addition to HOME, XDG dirs, and AKM_STASH_DIR.
• akm health is available for post-upgrade runtime checks against state.db, task-history integrity, agent availability, and recent improve telemetry.

Primary public command family for 0.8.0:

• akm improve [<type>|<ref>]
• akm propose <type> <name> (--task "..." | --file <path>)
• akm proposal list
• akm proposal show <id>
• akm proposal diff <id>
• akm proposal accept <id>
• akm proposal reject <id> --reason "..."

Proposal-queue and improvement workflows are consolidated around this family. Update scripts, prompts, docs, and agent instructions before upgrading.

Behavior changes worth noting:

• akm search with no query now fails with MISSING_REQUIRED_ARGUMENT.
• akm remember --enrich is fail-soft: if no LLM is configured, the memory is still written.
• akm wiki stash <wiki> <url> once again fetches and stores URL snapshots in wikis/<name>/raw/.
• akm health --since 24h provides a fast operator check for state-db round-trip health, task-log backing, and recent improve outcomes.
• The akm improve lock is now properly released even when openStateDatabase() or budget-timer setup throws, preventing lock leaks (up to 4 hours) on error paths.

New in the final 0.8.0 RC cycle:

End-of-Run Git Sync (akm improve --sync)

akm improve now auto-commits the git-backed primary stash at the end of every non-dry-run pass. Detection is based on a .git directory — no remote required.

• default and thorough profiles: sync + push enabled by default
• quick and memory-focus profiles: sync disabled (lightweight passes)
• --sync / --no-sync: force on or off for a single run
• --push / --no-push: push or commit-only override for a single run

The commit message defaults to akm improve auto-sync and supports {token} placeholders: {timestamp}, {date}, {time}, {scope}, {refs}, {accepted}. Configure via profiles.improve.<name>.sync.message in config.

The result envelope includes a sync field: { committed, pushed, skipped, reason? }. A stash_synced event is emitted to state.db. Sync failures are non-fatal — they never fail a successful improve run.

Consolidate Reliability

• Pre-flight stale-DB filter: consolidate now filters out entries that have become stale in the DB before making any LLM calls, eliminating wasted tokens on already-processed content.
• Defense-in-depth merge guards: three layers of guards prevent incorrect merges, improving consolidation correctness on edge cases.

Proposal Drain Refinements

• DrainDeferReason now has two variants: "mid-band" and "possible-dup". The "possible-contradiction" variant has been removed.
• DrainResult gains a staged field: items the judgment tier resolved as "accept" that were not yet promoted in queue mode. These are reported as resolved, not deferred.
• contentBodyLineCount now uses the canonical parseFrontmatter for consistent CRLF/BOM handling.

Workflow when_to_use Now Optional

Workflow frontmatter no longer requires when_to_use. The field is still supported and recommended but the validator no longer rejects workflows that omit it.

Release validation status for the final 0.8.0 cut:

• Full repo test suite passed.
• Docker install matrix passed for both Bun and binary install paths.