feat(verify)!: strict-by-default with --lax escape (ADR-0023 cross-port consistency) (#96)

[dmealing]

What & why

meta verify (TS) and metaobjects verify (Python) now load metadata strict by default. An undeclared or typo'd own @attr is ERR_UNKNOWN_ATTR and verify exits non-zero. This closes the ADR-0023 gap where such an attr silently passed verify in TS/Python but was rejected by Java's Maven metaobjects:verify goal (which already forces strict via LoaderOptions.create(false, false, true)).

A new --lax flag restores the previous open-attr load.

Closes #96.

Breaking change

verify is strict by default. A project whose metadata carries an attr no provider declares — a typo, or a vendor attr that was silently tolerated — now fails verify.

Scope: only verify defaults strict. gen / docs / agent-docs keep loading lax (a possible follow-up: extend strict to gen, deferred here).

Migration (the failure message names all three):

• register the attr on a metadata provider, or
• move arbitrary author-supplied properties into an attr.properties bag, or
• run meta verify --lax to keep the legacy open-attr load.

Before / after

# metadata with a typo'd @ttribute on a field.string

# before
$ meta verify        → exit 0  (silently passed)

# after
$ meta verify        → exit 1  ERR_UNKNOWN_ATTR + actionable hint
$ meta verify --lax  → exit 0  (legacy behavior, opt-in)

Implementation

• TS — LoadMemoryOptions.strict threaded into the loader; verify passes strict: !flags.lax; parseVerifyArgs gains --lax; the ERR_UNKNOWN_ATTR catch prints the three-exit hint.
• Python — _load_root(..., strict=False) + _generate(..., strict=False) threaded; --codegen / --templates verify paths pass strict = not args.lax; --lax added to the verify subparser; same actionable hint. gen / docs keep the default-lax load.
• Forge layer strict-clean (sdk) — loadMemory bundles Meta Forge descriptive types (decision/principle/…) + @forge* provenance attrs so mixed content loads. Under strict these were rejected, so the forge provider now admits its types under metadata.root and registers @forge* as common attrs. Also dropped the retired @dbTable no-op attrs from the trainer-website test fixture (source.rdb @table already holds the physical name).
• Cross-port fixture — fixtures/verify-strict-conformance/ (a registered node carrying one made-up @attr), asserted by both the TS and Python CLI verify-strict tests. The loader-level cross-port ERR_UNKNOWN_ATTR gate already exists at fixtures/conformance/error-unknown-attr/ (all five ports); Java's Maven goal already enforces strict.

Test evidence

RED → GREEN per port:

• TS verify-strict.test.ts — RED: strict default returned 0 / --lax unknown; GREEN after impl (3 pass). args-verify.test.ts updated for the lax field + a --lax parse test.
• Python test_cli_verify_strict.py — RED: --codegen/--templates returned 0, --lax unrecognized; GREEN after impl (5 pass).

Full suites:

• TS bun test — 4826 pass, 2 fail (both pre-existing, confirmed on the clean base: a metamodel-docs byte fixture + a package-manager-detection test; neither touches verify/strict/forge). Whole-workspace typecheck green.
• Python pytest tests/codegen — 257 pass, 2 fail (both pre-existing: test_cli_staleness_nudge asserts an old "metaobjects agent-docs" string vs the current "npx meta agent-docs" message — unrelated to strict load).
• Fixed the 2 real verify-subverbs failures my change first surfaced (the forge-layer + stale-fixture findings above).

🤖 Generated with Claude Code