feat: chunked, resumable dumps with breathing intervals + R2 overflow

[Furox-Art]

Closes #59

Problem

The legacy /export/dump path accumulated the entire database in a single in-memory string inside one request. On large databases this both blows the 30s Workers request window and exhausts memory, and a mid-request crash loses all progress.

Solution — bounded, resumable dump engine

New ChunkedDumpEngine (src/export/chunkedDump.ts):

• Rowid-batched table scans (rowsPerBatch, default 200) — bounded memory per batch.
• Time-budgeted cycles (cycleTimeBudgetMs, default 4s) — each cycle does bounded work, then yields with a breathing interval (breathingIntervalMs, default 5s) so concurrent queries are starved for at most one cycle, not the whole dump.
• Progress persisted in DO storage (tmp_dump_state) after every yield — a crashed/restarted dump resumes exactly where it stopped, and a resume never re-emits already-serialized schema (tableIndex advances before the yield point).
• R2 overflow: each chunk spills to R2_DUMP_BUCKET under <dumpId>/00000000.sql-style keys when bound; DO storage keeps the tail window as fallback, and reassembly is a streamed concatenation — the client never holds the whole dump in memory.
• Injection guard: table names are filtered through an identifier whitelist before being interpolated into any SQL.
• Backwards compatible: small DBs still get the old synchronous response; only when the 30s window is threatened does the route upgrade to a job.

Job flow (src/export/dump.ts, src/do.ts, src/handler.ts):

• POST /export/dump → synchronous response for fast dumps; otherwise returns { jobId } immediately.
• Continuation is driven by DO alarms (self-rescheduling cycles with breathing delays — no external cron).
• GET /export/dump/status (DO RPC dumpJobStatus) → progress table + downloadable stream when complete.

Spec checklist (issue #59)

• Chunked dump — rowid batches + byte-sized chunks (chunkTargetBytes)
• Breathing intervals between chunks (configurable, default 5s)
• Exceeds 30s → job continues via DO alarms, request returns immediately
• Large DB → R2 (>100MB friendly: chunked put + streamed concat read)
• No memory blow-up — nothing accumulates unbounded
• Status/poll endpoint for progress + download

Advantages over #76 / #93

Aspect	This PR
Resumability	Crash-safe state machine; resume tested (no duplicate DDL emission)
Memory	Bounded at every step: batch, chunk, and reassembly are all streamed
Yield fairness	Explicit breathing interval between cycles, driven by alarms
Storage	Dual-path: R2 when bound, DO storage fallback, streamed reassembly
Back-compat	Small dumps keep the old synchronous contract unchanged
Tests	12 unit tests: resume-mid-schema, resume-mid-data, chunk flush, R2 writes, reassembly order, injection guard, serialization

Testing

• npx vitest run src/export/chunkedDump.test.ts → 12 passed / 0 failed
• Full suite: 163 passed; the 4 failures in src/rls/index.test.ts and 11 typecheck errors (do.test.ts mocks, cache, cdc) reproduce on a clean checkout of main — pre-existing, untouched by this PR.
• Regression: src/export/dump.test.ts still passes with the new engine wired in.

```
Tests 12 passed (12)
Test Files 1 passed (1)
```

[Furox-Art]

⬆️ Upgrade: consolidated R2 multipart finalize + presigned download URLs (7eed275)

Building on the chunked engine, dumps now land in R2 as one object instead of N chunk keys:

What's new

• finalizeDump() (alarm-driven, outside the HTTP window): after completion, all chunk records are merged into dumps/<dumpId>/<fileName> via an R2 multipart upload. Time-budgeted per alarm invocation, so even a multi-GB consolidation progresses across several invocations instead of dying in one 30s window.
• Crash-resumable at the byte level: every uploaded part's partNumber+etag is persisted in DO storage; an interrupted finalize resumes with resumeMultipartUpload and continues from the exact byte offset — no part is ever re-uploaded (unit-tested).
• Presigned download URLs: /export/dump/status now returns downloadUrl (expiring, 1h) when the runtime exposes R2Bucket.createSignedUrl. Feature-detected — older bindings gracefully fall back to the existing streaming endpoint (also unit-tested for both shapes).
• Cleanup: per-chunk R2 mirrors are deleted after a successful complete (DO-storage records remain as the streaming fallback). assembleDump prefers the consolidated object.
• New tuning params: ?partBytes= / ?finalizeMs= (part size defaults to the R2 minimum of 5 MiB for non-final parts).

Evidence

npx vitest run src/export/chunkedDump.test.ts → 19 passed / 0 failed  (12 existing + 7 new)

New tests cover: part sizing & ordering, byte-offset resume, budget-exhaustion mid-upload then continue, no-R2 finalize, createSignedUrl presence/shape guards, and consolidated-object preference in assembleDump. Full suite: 170 passed; the 4 rls failures are pre-existing on main (documented in the PR description).