Hard-coded recommendation audit

MMML treats scientific defaults and operational recommendations as claims, not as decoration. CI runs scripts/audit_hardcoded_recommendations.py; high-risk language must have a nearby [evidence: <claim_id>] marker or say UNVERIFIED.

The initial audit found and classified these active heuristics:

Area	Previous presentation	Current state	Evidence
Joint-training optimizer selection	Recommended hyperparameters	Compatibility flag for an unverified legacy heuristic	[evidence: joint_optimizer_hyperparameters]
Hybrid heat ECHECK scaling	Recommended floor	User-overridable legacy heuristic	[evidence: cluster_echeck_scaling]
Pre-SD bonded-energy threshold	Safe default	Unverified heuristic threshold	[evidence: pre_sd_recovery_threshold]
JAX persistent cache	Safe default	Unverified environment policy	[evidence: jax_persistent_cache_policy]
DCM liquid shell values	Safe liquid defaults	Unverified recipe values	[evidence: dcm_liquid_shell_defaults]
TIP3 count-at-30-A recipe	gpu09-validated/pinned	Unverified historical recipe	[evidence: tip3_count30_recipe]
Solvent-burst matrix	Default validated matrix	Unverified campaign hypothesis	[evidence: solvent_burst_default_matrix]

The scanner intentionally excludes tests, build products, agent worktrees, and vendored Packmol sources. It is a guardrail for high-risk wording, not a proof that every scientific sentence is correct. The evidence registry remains the authoritative inventory, and reviewers should add claims when subtler language implies support, accuracy, stability, or transferability.

Cluster paths and checkpoint locations are not scientific defaults. Active workflow launchers now resolve them from MMML_CKPT, workflow configuration, or repository checkpoint resolution; examples use /path/to/... placeholders. Scheduler/node selections belong in versioned environment or workflow profiles.