fix(cli,importers): stop importer warnings from smearing the sync frame

[upsetbit]

The report

prosa sync came back with a garbled progress display:

2026/08/28 20:12:14 WARN codex: malf→ local        import→ local        importing  1308 / 1575 · imported 44 …

Two independent facts produce it.

1. The source data really is torn — the warnings are correct

Codex appends to sessions/**/rollout-*.jsonl while a session runs, and a torn
write occasionally leaves a line cut mid-token:

line 521  {"timestamp":"2026-08-28T19:43:50.796Z","ordinal":520,"type":"inter_agent_communicatio
line 522  {"timestamp":"2026-08-28T19:44:59.396Z","ordinal":520,"type":"inter_agent_communication_metadata","payload":{…}}

Measured across one day's tree: 17 torn lines in 64 041 lines, hitting 15 of
125 files. Every line number in the warning output points at a genuinely
invalid line, and errors 0 confirms the imports themselves succeeded. Not a
buffer problem either — the scanner is already at 16 MiB and the longest line
in the worst file is 1.3 MB. No parse decision changes in this PR.

2. Nothing owned the terminal

spinner.Run drives Bubble Tea on stdout with no alt-screen, so the
standard renderer repaints by moving the cursor up over its own last-frame line
count. Importers called package-level slog.Warn → slog.Default() →
stderr, unbuffered, while that frame was live. Each interleaved write
desyncs the renderer's line accounting.

The mechanism to prevent this already existed but covered only the pusher:
push.logger is swapped for warningCounterHandler and the tally surfaces as
the Warnings summary row. Importers had no equivalent hook.

The fix

ImportOptions gains a scoped Logger *slog.Logger, resolved through
importerutil.Logger(opts) (nil → slog.Default()). codex and claude-code
thread it into parseSession via a closure, so ParseFunc keeps its signature
and the four importers that never log stay untouched; grok-build passes it down
as a parameter because its Importer is a shared singleton. runSyncInteractive
points both the pusher and the importers at the same counter, so the existing
summary row reports both phases.

slog.SetDefault stays out of this path — #154 deliberately removed it, and
TestReconcileUsesPusherLogger locks that in.

Separately, malformed lines are now reported once per file (count + first
offending line + first error) instead of once per line, and codex/claude-code
skip blank lines instead of counting them as malformed. grok-build's own
emptiness guard is left byte-exact on purpose: the lines it collects are hashed
verbatim into the projection, and that hash is the dedup key, the sync hash, and
RawHash at once — TestImportHashUnchangedByBlankLines locks it.

Verified locally

Isolated store (scratch HOME), 67 real Codex files including all 15 damaged
ones, run under a pty so the interactive path actually engages.

	before	after
WARN lines inside the frame	17	0
summary	—	Warnings 15 diagnostic logs suppressed in TTY; use --verbose to see them
sessions / turns / session_tools / session_usage / kinds	66 / 2558 / 87 / 66 / 7	identical
sha256 of all session rows (id,raw_hash,raw_size,first_prompt)	4879c863…	4879c863…
sha256 of all turn rows	f850d8f7…	f850d8f7…

Import output is byte-for-byte unchanged. Also confirmed:

• --verbose emits 15 aggregated warnings summing to exactly 17 skipped
  lines, each carrying count= and first_line=.
• A re-run is a no-op (imported 0 · skipped 67) with zero warnings —
  RunSingleFile short-circuits on the hash before Parse.
• --json still keeps stdout pure NDJSON with diagnostics on stderr.

just ci and just quality pass.

🤖 Generated with Claude Code

https://claude.ai/code/session_01SGJVWRRZLmai3UHR7cHtJ6