Engine-version strata from `.translate/state`

[mmcky]

F1/F2 need stratification by engine version. Read tool-version / model from the state file history (not just HEAD) and stamp each ai-sync commit with the version in force; expose per-version totals in run.json. Depends on what the state files actually record (QuantEcon/project-translation#22 found model: unknown in 11 and 8 records and single-revision files in most of python.zh-cn/intro.zh-cn).

Part of QuantEcon/project-translation#49 (lifecycle metrics, Stage 1). Method and limits: docs/method.md.

[mmcky]

Delivered 2026-08-28 via #15. Every ai-sync commit in commits.jsonl now carries engine_model/engine_tool_version, resolved from the state-file history rather than HEAD — the record written by the commit or its wave (first revision at-or-after, within a two-day adjacency bound, exact for both the same-commit and adjacent-commit modes), otherwise the record last in force. run.json exposes per-version totals (engine_strata: sync commits, prose and character churn, overwritten prose by prior tier, ordered by first appearance) and overwrites.json entries carry the same stamp.

On the questions this issue flagged from QuantEcon/project-translation#22: the literal model: unknown records survive as their own stratum (one surfaced at python.zh-cn, unknown@0.20.0), distinct from unrecorded (no state file at the commit — pre-engine history), so the two cases stay separable downstream. Validated on the real corpora: programming.zh-cn resolves a ten-rung ladder from claude-sonnet-4-6@0.12.4 to claude-sonnet-5@0.26.0 with the 2026-04 wave isolated at 0.14.1; python.zh-cn splits the July resync wave (claude-sonnet-5@0.17.0, 69 sync commits, 1,134 of the edition's 1,249 human-overwritten prose lines) from every later version — the stratification the preserve-rule case in QuantEcon/action-translation#295 can now cite per engine version.