feat(fmlint): add param-fidelity rules to catch silent-discard params (X001-X003)

[marcoaperez]

FileMaker's fmxmlsnippet parser silently discards child elements whose tag doesn't match what a step expects — the step pastes fine, but the parameter quietly falls back to its default. These bugs are invisible at the XML-schema level and only manifest at runtime (a "Go to Record" that never advances, a Card window that opens as Document, an unresolved cross-file script reference).

What this adds

Three catalog-driven fmlint rules:

• X001 unknown-param-element — flags a child element not recognized by the step's catalog params[], with a difflib "did you mean" suggestion. Only enforced against catalog entries with status: "complete"; allowlists {Text, Animation} (elements FileMaker tolerates that the catalog deliberately doesn't model as params).
• X002 missing-discriminator — flags a param present without its required discriminator sibling (e.g. <Layout> without <LayoutDestination> on Go to Layout / New Window).
• X003 known-silent-discard-patterns — two hand-verified combinations: Style="Card" without the Styles bitmask (opens as Document instead); <FileReference> nested inside <Script> or missing <UniversalPathList> (cross-file script reference doesn't resolve).

Includes a synthetic test suite (11 cases) plus a mass smoke test against snippet_examples/ (0 false positives). The smoke test surfaced two catalog gaps, fixed here: the Layout param was missing entirely from New Window, and UniversalPathList was missing from Generate Response from Model.

Bonus fix

While porting this, found that the current step-catalog-en.json on main has invalid JSON — the Layout param entry under Go to Layout is missing a comma between "notes" and "discriminator", so any strict JSON parser (including this catalog loader) fails to load the file at all. Fixed in the same commit since the new tests need a loadable catalog to run.

Test plan

• python3 -m unittest agent.fmlint.tests.test_param_fidelity -v — 11/11 pass
• Smoke test against the full snippet_examples/ corpus — 0 false positives

[agentic-fm]

Really nice work, @marcoaperez — this catches a whole class of bugs (silently-discarded params) that are invisible at the XML-schema level and only bite at runtime. The difflib "did you mean" touch on X001 is a good call.

Verified end-to-end on my side: fetched your branch at f7a64f5, overlaid param_fidelity.py + tests + your catalog onto our internal tree, and ran the suite — 11/11 pass, including the corpus smoke test (zero X-family false positives across all of snippet_examples). The rules are correctly gated to status: complete catalog entries with forward-compat skips, and the New Window / Perform Script catalog additions line up with the rules that consume them.

One small thing before merge: since this was opened, main picked up the missing-comma fix (the same one your diff includes at ~L2958), so the catalog has moved under you. Could you rebase onto current main? The comma line will resolve to your side / as a no-op; the rest applies clean. Approving now so it's ready to go once rebased. 🙌