make real-time capture opt-in: default capture.realtime off, drop PostToolUse observe from adapters

[plind-junior]

what you're trying to do

capture should not run in real time by default. today every adapter ships a
PostToolUse hook that shells out to vouch capture observe after every
tool call, plus a Stop hook that shells out to vouch capture answer after
every turn. on a busy session that's hundreds of python process spawns, each
one loading the kb store, to append a line to an ephemeral buffer that only
matters once — at SessionEnd.

the knowledge itself is already batched: capture.answer_mode defaults to
session (capture.py:45), so claims are
extracted once at finalize from the full transcript, and the per-turn Stop
hook is a pure deferral — it fires, decides it is in session mode, and exits.
the real-time half of capture is doing work whose only consumer is a
once-per-session rollup.

there is also a correctness angle. the buffer is per-tool-call hearsay: a
Command failed: line, a truncated grep -rn "delete" src/vouch/*.py | head.
the transcript that finalize already reads is strictly richer and is the
receipt-bearing artifact. two sources of truth for "what happened in this
session", one of them lossy.

what you've tried

capture.enabled: false is the only existing knob and it is all-or-nothing —
it turns off the session summary too, which is the part worth keeping. there is
no way to say "capture, but only at the end".

suggested shape

1. new config key capture.realtime, default false. when off,
   capture observe returns a {"skipped": "realtime-disabled"} no-op
   immediately, before touching the store.
2. drop the PostToolUse block from the shipped adapter templates
   (adapters/claude-code/.claude/settings.json and the equivalent in the
   other host adapters), and drop Stop → capture answer while we're there
   — under the default answer_mode: session it can never file anything.
   SessionStart → finalize-all, UserPromptSubmit → context-hook, and
   SessionEnd → finalize are the three that carry weight.
3. reconstruct tool activity from the transcript inside finalize instead of
   from the buffer, so the summary body keeps its "files modified / activity /
   notable commands" sections without the per-call hook.
4. capture.realtime: true restores today's behaviour for anyone who wants
   the buffer as a crash-resistant backstop.

compatibility considerations

this is the part that needs care. session_split.summarize gates on
len(observations) + len(changed_files) < cfg.min_observations (default 3,
session_split.py:131). with the buffer
empty, only the git-diff backstop counts — so any session touching fewer than
3 files would file no summary at all. item 3 above is therefore not
optional; it is what keeps the gate meaningful. the alternative is counting
transcript exchanges toward total.

existing kbs are unaffected — no stored artifact changes shape. adapter
templates are regenerated by install_adapter, so installed hosts keep their
current settings.json until they re-run the installer; worth a line in the
changelog either way.

not a surface change, so no VEP: capture observe stays on the method list
and stays callable, it just defaults to a no-op.

alternatives

• keep the hook but make it write nothing when answer_mode: session — same
  saving in effect, but leaves the confusing "hook fires, does nothing"
  shape and still pays the process-spawn cost.
• debounce observe (write at most every N seconds). cheaper to build, but
  keeps two sources of truth and still spawns a process per tool call to
  decide it should not write.