feat: implement catalog and generator overhaul foundations

[kim-em]

Implements the LeanEval-side foundations from #536:

• replaces the legacy test flag with lifecycle, visibility, revision, and tag metadata across all manifests;
• adds catalog validation, deterministic v1 audit tooling, and freezes the mechanically selected 118-problem v1 set;
• publishes and pins the standalone lean-eval-generator JSON contract at exact commit a726789593eeac5c32ad82760061cd5bf6cae662;
• adds the external generator consumer seam while retaining the embedded implementation as a parity reference;
• adds the two reviewed software-verification drafts with the corrected statements, citations, and anti-vacuity guards.

The full v1 evidence is checked in under audits/v1/. The selection uses results commit 269c4dc9e3d264fe6b06e7d5d2fd1b0d86ac17e4, has 118 exact statement revisions, and has no manual additions or exclusions.

Verification:

• 25 Python catalog/audit tests;
• 94 generator tests;
• 19 module-coverage tests;
• structure-only manifest validation;
• targeted builds of both new Lean modules (only their intentional benchmark-hole sorry warnings);
• catalog validation reports 299 problems, 1 tag, and 1 frozen named set;
• exact published generator pin resolves and builds.

Tracked generated workspaces intentionally remain unchanged in this source PR. The trusted post-merge regenerator will refresh them against the new catalog and published generator revision.

[kim-em]

Fresh pre-merge v1 freeze audit: results main remains 269c4dc9e3d264fe6b06e7d5d2fd1b0d86ac17e4; the audit reads 44 files / 1,281 accepted records and reproduces the checked-in evidence byte-for-byte (336be9fcafa4730cd0daf4598d30ae10104c3f733a08aeef48ae2357a2518817). Applying the approved rule yields exactly 118 members, exactly equal to manifests/sets/v1.toml, with canonical membership digest 6d691218009d7afcea82e46c6de8bfa8a6d5de310e9b111899b24978f8fbf241. No additions, exclusions, or delta.