…eckpoint write (03-R1/06-M1)

The opencode plugin's parallel session_checkpoint `remember --force` write — the
sibling of the Claude-plugin write deleted in #80. It fired on every
stop/idle/compacted/deleted event + per-turn (every N positive touches) + before
improve/propose/evolve + a manual akm_memory action, flooding the stash with
below-every-gate write-only telemetry. Meta-review 03-R1/06-M1 retires it.

Deleted (opencode/index.ts): `captureSessionMemory` (the session_checkpoint
body-builder + `remember --force` write), `maybeCheckpointSessionMemory`,
`ensureFreshProposalCheckpoint`, and their helpers (formatEvidenceSummary,
formatPathBullet, countByValue, uniqueRecent, sessionHasPendingCheckpointEvidence,
getHarnessStatePaths) + dead state (AKM_MEMORY_CHECKPOINT_EVERY,
sessionFinalMemoryCaptured, sessionSuccessfulAssetTouchCount). Removed all call
sites: stop/idle handlers, the per-turn positive-feedback checkpoint, the
akm_memory `checkpoint` action (dropped from the enum too), and the pre-improve/
pre-propose/pre-evolve checkpoints. `akm index` on session-end survives —
`maybeIndexSessionMemory` now runs decoupled from the deleted write.

PRESERVED: the candidate → /akm-memory-promote pipeline
(extractCandidatesFromText/appendCandidates), the SessionEnd `akm extract` path,
and `rememberTextAsMemory` (kept — the akm-curator (akm_evolve) output write at
line ~3545 still uses it; that is a deliberate curator artifact, not the
session_checkpoint telemetry, and is out of 03-R1's scope).

Tests: deleted 5 that asserted the checkpoint/stop write + migrated 1 to assert
`akm index` still runs on stop with NO `remember` call. Gate: `bun test tests/`
314/0, `tsc --noEmit` clean, `tier2/runner.ts --baseline` exit 0. Net −542 LOC.

(Done directly, in Opus — the opencode/gpt-5.4 agent stalled on an OpenCode Zen
monthly usage-limit; small requests trickled through but the deletion task hung.)

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018YjU1mGiShdDP4qFW7p4mv

…eted with 03-R1)

The prior commit deleted `captureSessionMemory`, which did TWO things: the
session_checkpoint `remember --force` write (correctly retired by 03-R1/06-M1)
AND the memory-candidate extraction that feeds the owner-gated
candidate -> /akm-memory-promote pipeline (a KEEP reference rung). The second
went out with the first, so no candidates were harvested on session end.

This passed `bun test tests/` locally only because three candidate tests read a
persistent on-disk candidate log seeded by earlier runs under the same stable
session IDs; CI's clean checkout had no such state and the tests failed
(`expect(candidates.length).toBeGreaterThan(0)`).

Fix: add `maybeExtractSessionCandidates(sessionID, reason)` — harvests explicit
"remember ..." intents from the session buffer into the candidate log with NO
stash write, wired into `stop` (the reliable terminal flush) and `session.deleted`
(salvage before the buffer is discarded). Deleting the buffer after extraction is
the de-dup that the removed `sessionFinalMemoryCaptured` Set used to provide.
Candidate text is now sourced from memory-intent notes only (higher signal than
the old full-checkpoint-body scan), which suits 03-R1's noise-reduction intent.

Not wired into the recurring `session.idle`/`session.compacted` events — those
fire per-turn and a per-turn buffer flush is exactly the churn 03-R1 retires;
`session.idle` already runs the `akm extract` backstop.

Gate (clean isolated HOME/XDG, CI-equivalent): tsc 0 errors, `bun test tests/`
314/0.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018YjU1mGiShdDP4qFW7p4mv