fix(seed,import): sequential compatibility writes bypass transient retry

[yinlianghui]

Baseline and relation to #2678/#2680

• Verified at 98874656ffc50ce1531af52346228ffcdda73fba.
• perf/reliability: seed + data-import bypass the engine's bulk-insert path → N round-trips, N summary recomputes, no transient retry #2678 requires transient seed/import write errors to be retried.
• perf(seed,import): route bulk writes through the engine's batch insert path #2680 added withTransientRetry, but deliberately preserved two sequential compatibility paths without wrapping their writes: self-referencing seed datasets and import protocols without createManyData.
• This is an incomplete acceptance boundary, not a regression from the pre-perf(seed,import): route bulk writes through the engine's batch insert path #2680 sequential behavior.

Minimal reproductions

1. Real SQL self-referencing seed dataset (employee.manager) with one injected fetch failed on Alice insert, Bob insert, or update mode: each target operation is attempted once.
2. Import protocol exposing createData but no createManyData, with one injected fetch failed: the create is attempted once and reported failed.

Each scenario reproduced 2/2 on fresh state.

Expected / actual

• Expected: preserving sequential ordering/compatibility does not exempt the individual write from transient retry.
• Actual: SeedLoaderService.writeRecord and the import fallback call the engine/protocol directly, bypassing withTransientRetry.

Impact

Self-referencing seeds and compatible third-party/minimal protocols can still drop rows on a one-off network error—the exact reliability failure #2678 set out to remove.

Evidence and exclusions

• Seed source: packages/metadata-protocol/src/seed-loader.ts self-ref branch → writeRecord.
• Import source: packages/rest/src/import-runner.ts no-createManyData inline fallback.
• Normal bulk seed/import paths and import updates passed transient retry QA.
• Sequential ordering and the no-bulk protocol fallback should remain; only retry coverage is missing.

Required regression tests

• Self-ref insert/update transient failure retries and preserves reference order.
• No-createManyData import transient create retries without changing per-row compatibility semantics.
• Persistent and logical failures retain bounded attempts and row-level errors.

[yinlianghui]

可复现性审计补充（2026-07-11）

固定基线仍为 framework 98874656ffc50ce1531af52346228ffcdda73fba（归因基线 21420d9f82ebdcd53a6361ded3d829723bcab18e）。

两条路径的独立 fixture 与判定

1. 自引用 seed 顺序路径

   • 对象：employee；name 为自然键，manager_id 为指向 employee、按 name 解析的 lookup。
   • 数据：Alice，Bob.manager_id=Alice；另以已有记录运行 update 模式。
   • 故障：分别在 Alice insert、Bob insert、update 的目标单条写入第一次抛出 Error('fetch failed')，第二次应成功。
   • 成功标准：每个目标写入均尝试 2 次，引用顺序保持正确，最终无行丢失。
   • 当前结果：每个目标写入仅尝试 1 次并失败；三种注入点均为 2/2。

2. 无 createManyData 的 import 兼容路径

   • protocol 只实现 findData/createData/updateData；输入一行 { name: 'R1' }。
   • createData 第一次抛出 Error('fetch failed')，第二次应返回 { id: 'id_R1' }。
   • 成功标准：createData 调用 2 次，summary 为 created=1/errors=0。
   • 当前结果：createData 仅调用 1 次，summary 为 created=0/errors=1；2/2。

[yinlianghui]

暂时关闭，等待人工逐项复核。当前 QA 复现证据保留；本次关闭不表示问题已修复、无效或重复。人工确认后将按需重新打开。