feat(capture): make realtime observe opt-in (default off)

[kurosawareiji7007-hub]

Summary

• Add capture.realtime (default false): capture observe no-ops with {skipped: realtime-disabled} instead of spawning a store write on every tool call.
• Drop Claude Code adapter PostToolUse / Stop hooks; SessionStart/UserPromptSubmit/SessionEnd remain.
• Rebuild observations from the transcript at finalize so min_observations still works without the realtime buffer; realtime: true restores the old buffer path.
• Host-neutral rebuild: try Claude parser then Codex (same as load_transcript), including Codex tool mapping via _observation_from_call.
• CLI checks enabled: false before realtime so a fully-disabled kb stays silent.
• Installer prunes vouch-owned PostToolUse/Stop observe/answer hooks on reinstall so upgrades drop per-tool spawns.
• Note: rebuild inherits parser max_messages=2000; very long sessions may lose oldest tool calls vs uncapped realtime buffer.
• Test stub rewrites (sh→python) are intentional Windows-compat drive-bys alongside yaml config injection.

Fixes #602

Test plan

• pytest tests/test_capture.py tests/test_session_split.py tests/test_adopt.py tests/test_capture_answer.py tests/test_session_transcript.py tests/test_install_adapter.py
• Codex transcript rebuild case in test_observations_from_transcript_rebuilds_codex_tool_activity
• Silent when enabled: false even with realtime: false
• Installer upgrade prunes old PostToolUse/Stop vouch hooks
• Confirm starter config ships realtime: false
• With default config, vouch capture observe emits skip JSON and does not write a buffer
• Finalize with a transcript still files a session summary when activity clears min_observations
• capture.realtime: true restores PostToolUse-style buffer observe

[plind-junior]

@kurosawareiji7007-hub Check the CI failing issues

[plind-junior]

thanks — this lands all four items of #602's suggested shape: the capture.realtime knob defaulting off, the early no-op with {"skipped": "realtime-disabled"}, dropping both PostToolUse and Stop from the shipped template, and rebuilding observations from the transcript so the min_observations gate stays meaningful. the leftover-buffer fallbacks in session_split.summarize and the regression coverage are exactly right.

one change is needed before merge. observations_from_transcript (src/vouch/capture.py) imports parse_claude_transcript directly, but the default flips for every host. a codex session that passes its transcript to finalize rebuilds zero observations, falls back to an empty buffer, and any session touching fewer than min_observations files silently stops producing summaries — the exact failure the issue's compat section flagged, now on non-claude hosts. please update the rebuild to reuse the format detection already in transcript.load_transcript (try the claude parser, fall back to parse_codex_transcript) instead of hardcoding the claude one, and add a codex-transcript case to test_observations_from_transcript_rebuilds_tool_activity.

two smaller updates worth making in the same pass. in capture_observe_cmd (src/vouch/cli.py), swap the order of the two config checks so enabled: false stays silent regardless of realtime — right now a fully-disabled kb prints the skip json where it used to print nothing, and the existing silent-when-disabled test only covers enabled: false, realtime: true. and in the settings merge, prune vouch-owned PostToolUse/Stop entries (match on vouch capture observe / vouch capture answer) so re-running the installer actually removes the per-tool spawns for existing installs; the current assert "PostToolUse" not in merged["hooks"] in tests/test_install_adapter.py passes trivially because the fixture never had the old hooks, so please add that upgrade case too.

two optional notes, no change required: the rebuild inherits the parser's max_messages=2000 default, so very long sessions lose their tail tool calls where the buffer had no cap — worth a higher cap or a comment acknowledging the bound. and the sh→python stub rewrites in the test files are windows-compat drive-bys; the yaml conversions are justified by the config injection, just call the rest out in the pr body so reviewers know they're intentional.

happy to re-review once the host-neutral rebuild is in — everything else here is in good shape.

[plind-junior]

Check the comment above

[kurosawareiji7007-hub]

Addressed the review items on 312df0c:

1. Host-neutral rebuild — observations_from_transcript now tries Claude then Codex (same detection as load_transcript), maps Codex tools via _observation_from_call (unwraps the parser's arguments/input wrapper), and adds a Codex case to the rebuild tests.
2. CLI order — enabled: false returns silently before the realtime skip JSON; covered by test_observe_cli_silent_when_disabled_even_if_realtime_off.
3. Install prune — _merge_settings prunes vouch-owned PostToolUse/Stop hooks matching vouch capture observe / vouch capture answer; upgrade case in test_settings_json_merge_prunes_retired_observe_hooks.

Also noted the max_messages=2000 bound in the rebuild docstring and called out the Windows test stub rewrites in the PR body.

[kurosawareiji7007-hub]

CI is green on 35d74bc (including the 100% diff-coverage gate). Review items from the earlier pass are in; ready for re-review when you have a moment.

[plind-junior]

LGTM