fix(sessions): settle native dedupe ties deterministically

[shabhui]

Problem

#1383: NativePiSessionService.list() dedupes copied session files by keeping the newest write per native id, but updatedAt is derived from file content. A byte-identical copy always ties with the original, so "keep the newest write" never resolves the exact case the dedup (#1359 / #1374) was added for. The tie fell through to walkJsonl order, which follows directory enumeration and differs per platform:

• NTFS (alphabetical readdir + LIFO stack): the backup/ copy is walked first and wins, the original's record is deleted from records, and the new regression test from fix(sessions): specify native duplicate identity handling #1374 fails on Windows.
• ext4 (hash order): the original happens to win, so CI stayed green and masked the difference.

Consequences on Windows: a kept .jsonl backup takes the session over — reopening from the sidebar appends to the copy while the original freezes, persisted references to the original id get NOT_FOUND, and a live runtime's record can be deleted mid-session.

Fix

Order dedupe ties deterministically instead of trusting enumeration order:

1. an id with an open runtime keeps the file that runtime is already writing;
2. the last tie is settled by path order, which every platform computes the same way.

Tests

• breaks copy ties by path order instead of directory enumeration (#1383) — a copy in a directory sorting before the original's group still yields exactly one entry, and the winner is deterministic across platforms.
• keeps the open runtime's file when a copy ties on recency (#1383) — after a faux-model turn, a byte-identical copy sorting before the original loses to the open runtime.
• The existing #1359 test (copy in backup/, original wins) now passes deterministically on all platforms; it failed on NTFS before this change.

Local validation (Windows 11 / NTFS):

• vitest run src/native-pi-session.test.ts — 39/39 pass (37 before + 2 new)
• tsc -b packages/agent-runtime — clean
• biome: packages/agent-runtime is outside the repo's biome.json include list (checked, no findings possible)

fixes #1383

[vastsa]

Maintainer follow-up (a2eb0a5, included in this PR)

1. The [Bug] 会话列表出现重复会话：递归扫描 + 按路径生成 ID，子目录副本变成第二条会话 #1359 assertion was a no-op. That test built its "control" scan after the copy existed, so it compared the winner with itself and passed on every platform — including the NTFS/APFS enumeration order where the copy wins. Reproduced on APFS against the original code: the copy takes the id over, and the test still passed. So the "it failed on NTFS before this change" claim does not hold. The control now scans before the copy is written, which fails on main (copy wins) and passes with this fix.
2. Extracted the inline faux model runtime that two tests duplicated into fauxModelRuntime().
3. Documented the rankTie comparator (why it is a total order, and what the id fallback is for).

The tie-break logic itself is unchanged (open runtime → smallest record path) and was re-verified as order-independent under forced reversed readdir. Suite: 1240/1240, tsc --noEmit clean.

[shabhui]

感谢审阅、强化与合并!

补一个事后核对,两处更正:

1. 我在 [Bug] 会话去重(#1374)平局裁决依赖文件系统遍历顺序:Windows 上副本反向胜出,新回归测试在 NTFS 必挂 #1383 里说原 [Bug] 会话列表出现重复会话：递归扫描 + 按路径生成 ID，子目录副本变成第二条会话 #1359 测试"在 NTFS 上必挂",这个说法不成立。 原测试的 control(list() 扫描)发生在写入副本之后,胜者是在和自身比较,因此在任何平台上都能通过——它其实是个 no-op。您把 control 移到写入副本之前并抽出 fauxModelRuntime(),才让它真正具备了捕获能力(16950ee)。

2. 为了确认这一点,我在 Windows/NTFS 上把强化后的测试放回合并前的代码(831b66d)跑了一遍:2 failed | 37 passed——失败的正是强化后的 [Bug] 会话列表出现重复会话：递归扫描 + 按路径生成 ID，子目录副本变成第二条会话 #1359 用例和新增的 [Bug] 会话去重(#1374)平局裁决依赖文件系统遍历顺序:Windows 上副本反向胜出,新回归测试在 NTFS 必挂 #1383 路径序用例,与您在 APFS 上的结果一致。也再次印证 rankTie 的路径字典序裁决在 NTFS 上行为相同(字节级副本平局 → 副本胜)。

感谢修正,学到了:control 必须发生在被测变更之前,否则断言恒真。