fix(host-core): keep intact history readable around torn UTF-8 records

[AR307]

Summary

Follow up on merged #1565 with a focused host-side fix for the UTF-8 failure discussed
in #1560. The outbox change prevents one refused append from blocking the
queue; this change keeps intact transcript records readable when an interrupted
append leaves an incomplete multibyte character. It is independently mergeable
and does not alter the outbox retry or quarantine policy.

Reproduced failure

The transcript readers use BufRead::read_line(&mut String). An invalid UTF-8
sequence fails during that read, before the existing malformed-record handling
can run. A torn Chinese/emoji record can therefore hide otherwise intact
history, including valid records appended after the damaged line.

All five added regression tests fail on unmodified main 22560e083aa7 with
stream did not contain valid UTF-8, and pass with this change. This establishes
the torn-record defect; it does not establish every reported occurrence's
origin or prove the proposed connection with #1515. ends_with_newline already
reads a raw byte and is left unchanged.

Changes

• Find JSONL boundaries using bytes, then decode each record strictly.
  Log invalid UTF-8 without logging its content and skip that record on reads.
• Advance offsets over complete invalid lines, but leave an unterminated tail
  eligible for refresh when the writer supplies more bytes.
• Apply the same boundary to full, paged, indexed, tool-lookup and revision reads.
• Preserve unrelated raw bytes and CRLF during metadata updates instead of
  requiring the entire file to decode as UTF-8.
• Add focused regressions and extend the existing real-Electron regenerate/quit
  scenario to reopen a torn Unicode transcript before regenerating.

Reading does not truncate the transcript, replace characters, reconstruct lost
records or change the storage format. Invalid records stay on disk but are not
displayed. Existing explicit full-history rewrite behavior is unchanged.

Validation

Windows; existing Node 24 and Rust 1.90 toolchains; isolated app data and a
controlled local HTTP/SSE provider, without production credentials.

• Rust transcript suite: 42 passed, including the five new Unicode tests.
• Regression proof: the same five new tests against original main: 5 failed
  with invalid-UTF-8 errors, not fixture/build failures.
• Adjacent Desktop persistence, regenerate, shutdown and subagent-settlement
  regressions: 28 passed, including the six new fix(desktop): quarantine outbox appends the host keeps refusing #1565 cases.
• Workspace package builds, Desktop/pi-host typechecks, Electron
  main/preload/renderer build, fresh sidecar bundle and release Rust host build:
  passed.
• Source lint, style-token and agent-policy checks, cargo fmt --check,
  cargo clippy -p host-core --all-targets --release --locked, and diff checks:
  passed.
• node scripts/e2e-regenerate-quit.mjs --torn-unicode: 8 passed in the
  actual isolated Electron application. This covers a real Read tool response,
  reopening the same damaged session without a fork/model replay/file mutation,
  regenerate, native quit, restart and both revision branches. The resulting
  history screenshot was also inspected.
• Full Rust comparison: candidate 809 passed, 4 failed, 1 ignored; unmodified
  main 804 passed, the same 4 failed, 1 ignored. Failures are the two existing
  config-sync engine cases and the two data-relocation capability-override
  cases. The full suite is not reported as green.
• Rebased on main 06c04e99dc84 after fix(desktop): quarantine outbox appends the host keeps refusing #1565 merged and reran the combined
  Desktop build, adjacent regressions and actual Electron scenario.

macOS/Linux execution and the full Desktop/runtime test suites were not rerun
for this focused Rust follow-up.

Related to #1560; complements #1565.