[G1] Full-sync and hydration throughput baseline, memory ceiling, and a CI regression gate #176

Part of #160 (group 1 — independent).

Problem

docs/benchmarks.md covers shallow discovery and warm reads only (cold discovery 20/100/1000 = 9.19/10.29/47.67 ms; unchanged = zero files opened; warm session events 20/200 = 0.61/1.53 ms; OpenCode 1k vs 10k sessions flat at ~3.3 ms). There is no published number for full sync or full hydration throughput (records/s, MB/s), no memory ceiling, and scripts/benchmark-hydration.mjs has no committed baseline table. crates/ai-hist-cli/tests/discovery_bench.rs is #[ignore]d.

Once burn depends on relayhistory, burn's ingest --watch (1 s ticks) and burn summary --ingest inherit relayhistory's ingest cost. burn today parses a 100 MB Claude transcript incrementally in well under a second per tick; relayhistory must be measured before it can be trusted with that path, and the parity issues in this epic (markers, tool-result facts, incremental cursors) all change the hot loop.

Scope

1. Generator: scripts/gen-synthetic-history.mjs (or a Rust #[ignore] test helper) that fabricates deterministic Claude/Codex/OpenCode/Cursor/Grok stores of configurable size from the fixture corpus (e.g. N sessions × M turns × K tool results of B bytes), with a seed, so numbers are reproducible and nothing real is committed.
2. Benchmarks (scripts/benchmark-sync.mjs + Rust criterion or #[ignore] tests): cold sync --local on {100 MB, 1 GB} stores; incremental sync after appending 1 KB to one transcript; hydrate_session on a {1 MB, 50 MB, 200 MB} transcript, cold and unchanged; watch tick with nothing changed; report records/s, MB/s, files opened, bytes read, peak RSS, DB growth, WAL size.
3. Baseline table committed to docs/benchmarks.md with machine spec, date, commit, and the exact commands (match the existing table's format).
4. CI gate: a fast subset (≤ 60 s) runs in ci.yml on every PR against a small synthetic store and fails on > 2× regression in records/s or > 1.5× in peak RSS vs. numbers stored in scripts/benchmark-thresholds.json; the full matrix runs on a manual workflow_dispatch. Query-plan assertions stay the primary gate for reads (per current docs); this adds the write path.
5. Record the current numbers before the parity issues land, so their PRs show before/after.

Acceptance

• docs/benchmarks.md has a "Sync and hydration" table with the 8 rows above.
• The CI job exists, is green on main, and a deliberate 3× slowdown patch (test in a draft PR) turns it red.
• Peak-RSS measurement works on Linux and macOS runners (document how it is measured).

Out of scope

Optimizing anything (the incremental-hydration issue does that); OpenCode DB-copy removal (OpenCode parity issue).