Claims tokenize .dag text through tokenize_prepared(dag_prepared_lex())

[gunbai-bot]

Follow-up to #13294. 40 claim/support modules that tokenized .dag text with the dag rules (tokenize(rules: dag_lex()) or rules: lm.lex over dag_language_model()) now call tokenize_prepared(prepared: dag_prepared_lex()), so lex_compile_rules stops running per claim (#13294 floor: producer=lex_compile_rules claims=26 evals=26).

Left as-is: lexer-subject claims (test/claim/tokenize/*, self_host_contract/tokenize_emitted_contract), non-dag language claims, and production callers. Assertions unchanged.

Receipt (eval-step / CPU delta, cross-claim-demand line): to be posted from this PR's floor run.

🤖 Generated with Claude Code

[gunbai-bot]

On review 76613's one open point (fold_lowering_test.dag and the ambient name pool): the floor on run 37321193353 compiled and passed with that file switched to tokenize_prepared(..., prepared: dag_prepared_lex()), so both names resolve there. The current head is being re-checked by run 37328482005.

— sent from vivid-fox-372

[gunbai-bot]

Receipt from floor run 37340049841 on head cff2f6f, compared with #13294 floor run 37268749188:

• [cross-claim-demand] lines for prepare_lex_rules / prepare_lex_rule_set / lex_compile_rules (claims=26, 517 / 510 / 365 ms total, ~497 ms cross-claim): gone. None of the three appears in the new run.
• dag_prepared_lex shared by declared demand: 19 claims / 2 sites -> 265 claims / 47 sites, one fill (7 ms fill, 30 ms inclusive).
• prepare_lex_rules unadmissible (ArgumentNotClosedConstant): 39 claims / 4 sites -> 35 / 4. These four sites are production folds that prepare lm.lex (reference_conservation, reference_conservation_census, 00_compile, program_assembly / closure_emission), reached by claims. The brief excludes production callers, so I left them. They are no longer a cross-claim-demand row.
• CPU delta: about -0.5 s of per-claim lex preparation (the inclusive prepare_lex_rules cross-claim figure). Eval steps: the floor log does not attribute eval steps per producer, and the two runs select different claim sets (698 vs 774), so I have no clean whole-run step delta. The 14 touched claims that report shared-fill cost in both runs went from 17,835 to 14,529 measured CPU ms and from 4,437,775 to 4,349,149 measured eval steps.

Also switched, beyond the original 26 call sites: two claim sites, plus one new claim added on main, that re-prepared dag_language_model().lex inline. Assertions are unchanged throughout.

— sent from vivid-fox-372