[kvlonge]

Why

Custom rules spent most of their time preparing data rather than evaluating rules: the main process built Python records for the whole project and pickled them into one payload before the host could start (about 3 s of main-thread work on the dense 10k project).

Changes

• The custom-rule host payload is written natively as one file: model names, paths, keys, dependencies, plain config values, authored and expanded SQL, inferred and declared columns are stored without building Python objects; remaining sections are pickled per value.
• The host reads the file before the hermeticity guard, parses only the header and decodes each fact lazily on first read. Fact views read through small protocols, so unread sections are never decoded.
• Read tracking, rule cache keys, digests and evidence stay in the main process and are unchanged.
• AGENTS.md: performance measurement lessons (instruction counts under load, verifying the built binary, critical paths of concurrent phases, verify-or-discard speculation).

Verification

• Fact identity tests: every public fact decoded from the payload equals the full project value, for native and hand-built records.
• New CLI tests (every-fact rule with cold/warm/edit cache counts, failing, crashing and slow rules) pass unchanged against the previous code.
• Focused rule engine, rules, prepared-artifact and custom-rule E2E tests pass; 1k output identity holds.
• Dense 10k, 4 paired cold runs: custom_rules_ms 7.3-9.5 s to 4.0-4.6 s, CPU 58.8-60.0 s to 56.3-56.7 s; wall gain is smaller because built-in rules now bound the rules phase.
• Independent review: no blockers.
• make check-ci passes.