Search within the session log viewer

[jongracecox]

Reading a transcript back gave you no way to find anything in it, and the footer
advertised the home screen's keys regardless of the view.

What's new

/ searches the open transcript. Matches are highlighted in place, n/N
step through them with wraparound, and Ctrl+R/Ctrl+I toggle regex and case
as they do in the s transcript search. A match inside a folded tool call opens
that fold — a highlight you can't scroll to is worse than none.

Searching never rebuilds the view. _render_transcript now returns the
widgets plus a LogBody per body (widget, source text, the folds hiding it);
a query only calls Static.update() on the bodies whose spans actually changed.
Matching itself runs in a worker thread and returns plain span data — no widgets
are built off-thread — guarded by a generation token, since a scan can land
after the query has moved on.

Measured on a real 838-message log:

	before	after
per query	~540 ms, full remount	~1.7 ms (thread) + ~6 ms (event loop)
initial open	~190 ms	~140 ms, once

The remount was visible as the query box disappearing mid-keystroke.

The messages moved into their own scroll pane, so the header and query box
stay put while the transcript scrolls. The viewer's hint row is gone: / and
Escape are in the footer already, and a second strip of grey text under a
screenful of transcript reads as a second footer.

Backing out of a search result now lands on the result you opened rather
than the top of the list, so a set of results can be walked one at a time.

The footer only shows keys that do something here. check_action gates
them; / is two bindings on one key so it can read Filter on the home screen
and Search in the viewer, and the mode chords are no longer hidden now that
they only appear where they apply.

Two crashes fixed on the way

• A Collapsible title built from transcript bytes was parsed as console
  markup, so a Bash command containing [ raised MarkupError and took the
  viewer down. Titles, the viewer header and the session picker rows are now
  Content. Same reasoning for the query box's select_on_focus, which was
  eating the query whenever focus returned to it.
• A transcript the parser can't structure now shows a warning banner and the raw
  log — searchable like any other — instead of an error screen. A file that
  can't be read is still an error. iter_hits swallows the same class of
  failure per log, so one bad transcript can't end a scan over hundreds.

Instrumentation

Both swallowing paths are the ones --debug most needs to see, since neither is
visible in the UI: claude.transcript_unparsed carries the traceback and
search.scan_failed names the log it dropped. The viewer logs tui.log_view
and log_search.start/.done/.dropped, whose repainted count is what shows
the in-place update still working. Verified under --debug-redacted against a
real transcript: no username leak, query redacted.

Also

Documents that a global install needs --no-cache. uv's build cache is keyed on
the derived version, which hatch-vcs makes identical across untagged commits,
so --force --reinstall reports success while leaving the old wheel in place —
which looks exactly like a feature that doesn't work.

Testing

977 tests pass, coverage on cli.py holds at 95%, pre-commit clean.