Add rigorous free local language model screening

[DevHabito]

What changed

• records the Qwen2.5 1.5B artifact, load, timeout, and throughput diagnosis;
• records the free Qwen3.5 0.8B artifact, desktop load, benchmark, and live screens;
• adds a local-only five-level numeric grammar for UNDERSTAND while preserving the shared E044 schema;
• rejects unregistered local numeric levels without rounding, coercion, or retry;
• adds an admitted human-gated measurement runner that cannot send another registered prompt after a decisive failure;
• preserves the invalid E054 runner execution separately from the valid E055 model-quality failure; and
• updates the public English documentation to match the observed state.

Why

The 1.5B candidate was too slow on the target notebook. The faster 0.8B candidate exposed two separate issues: llama.cpp does not enforce continuous number bounds in its JSON-schema grammar subset, and the model can copy a stale expression instead of answering the current request. The change repairs only the demonstrated numeric grammar gap and records the remaining quality failure without promoting the model.

User and developer impact

• no paid provider or provider key is required;
• no automatic backend fallback is introduced;
• the language model still has zero authority over memory, goals, identity, RZS, sigma, tools, or actions;
• temporary conversation context is erased on close; and
• no local model has been promoted as conversationally adequate.

Validation

• E054 focused suite: 27/27 passed;
• E054 full suite: 500 run, 499 passed, 1 declared Windows symlink-privilege skip;
• E055 focused runner and local-boundary suite: 32/32 passed;
• E055 full suite: 505 run, 504 passed, 1 declared Windows symlink-privilege skip;
• E055 live screen: two gateway-valid pairs, zero authority mutations, zero provider cost, then a registered early stop on an exact stale-expression copy.