fix(sync): bound himalaya read concurrency and keep partial sync progress

[vkehfdl1]

Closes #33.

Root cause

HimalayaSource.list() read every envelope of a page with Promise.all, so the
default --page-size 1000 fired up to 1000 simultaneous himalaya message read
processes (one IMAP login each). Gmail throttles that fan-out, one read failed,
the whole page rejected, and because the CLI only called archive.sync() after
list() resolved, the archive was left untouched: no retry, no partial
progress, no usable chunk store.

What changed

• Bounded read concurrency (src/source.ts) — a page is read through a pool
  of at most --concurrency <n> processes, default 4, decoupled from
  --page-size, which stays the envelope window. A page whose read fails also
  stops dispatching new reads instead of spawning the rest of the window.
• Per-message retry with backoff (src/source.ts) — reads run in rounds:
  the first round reads the page, later rounds retry only the messages that
  failed, after a doubling backoff (250ms, then 500ms), three attempts per
  message.
• Partial progress + diagnostics (src/source.ts, src/cli/index.ts) —
  sources expose collect() returning { messages, failures }. sync commits
  the readable messages and reports failures[] (providerKey, attempts, and
  the redacted error); it exits nonzero only when nothing could be read.
• himalaya stderr preserved in the surfaced error, so throttling is
  distinguishable from a malformed message.
• Docs — README, docs/architecture.md, and the agent skill document the
  pool, the envelope window, failures[], and the exit rule.

Verification

• npm run typecheck, npm run build, npm test (24 files / 80 tests),
  npm run validate:skill — all green, including the existing fixture sync and
  status/embed/repair/doctor CLI contracts.
• New tests/issue-33-read-concurrency.test.ts (11 tests): pool bound and order
  preservation, dispatch stop after a failure, retry with growing backoff,
  partial progress, stderr preservation, the CLI JSON contract, the
  nothing-read exit rule, the empty-page boundary, and --concurrency
  validation.
• The throttle is reproduced with a stub himalaya binary driven through the
  real execFile + PATH spawn path (POSIX-only tests; Windows CI runs the
  in-process pool, retry, and collect coverage, and the CLI validation case).
• Each new assertion was shown to fail for the right reason before the fix
  (page abort on throttle, first-attempt rejection, sync abort with one
  unreadable message, dropped stderr), and mutation proofs confirm the new
  assertions can fail (unbounded pool, no backoff, retry disabled, failures
  dropped from the sync JSON, nothing-read rule disabled, empty page treated as
  a failure).

Real-account check

Read-only run on this machine against a real Gmail INBOX with the issue's own
repro (sync --page-size 25 --concurrency 4, temp data dir):

• The first attempt read the whole 25-message page at concurrency 4 and reached
  archive.sync(), then stopped on a missing Korean-analyzer shell export
  (MAILCRAWL_KIWI_MODEL) — an environment prerequisite from
  docs/multilingual-installation.md, unrelated to this change.
• Re-run with the documented analyzer exports (plus the JA/ZH helper paths) —
  exit 0, added: 25, failures: [], and status --json showing 25 messages /
  833 chunks / healthy FTS. Details in the PR comment below.

[vkehfdl1]

Real-account verification (post-fix)

Read-only run on this machine against the reporter's Gmail INBOX with the
issue's own repro, using the built CLI from this branch:

node dist/cli/index.js --data-dir <tmp> sync --source himalaya \
  --account gmail --mailbox INBOX --page-size 25 --concurrency 4 --json

Sync JSON (exit 0):

{"added":25,"updated":0,"deleted":0,"unchanged":0,"touchedThreads":25,"rebuiltThreads":25,"chunksAdded":833,"chunksDeleted":0,"embeddingBacklog":833,"excluded":0,"excludedByReason":{},"failures":[]}

status --json afterwards: messageCount: 25, chunkCount: 833,
fts: {status: "healthy", rows: 833}.

Before the fix the same command aborted after ~12s with
himalaya message read failed: … FETCH returned no body for the requested message and left messageCount 0 (reported in #33). No read failed this time.

Two notes on the run:

• The first post-fix attempt read the whole 25-message page at concurrency 4 and
  reached archive.sync(), then stopped on a missing MAILCRAWL_KIWI_MODEL
  shell export — an environment prerequisite from
  docs/multilingual-installation.md, unrelated to this change. The result
  above is from the re-run with the documented analyzer exports (and the JA/ZH
  helper paths) set.
• The temp archive created by this run was deleted afterwards; no message
  content is included here.