SIGN IN SIGN UP

fix(cache): register the percent flag everywhere, tell users when their budget is too small

Flag plumbing the first pass missed:
- --block-disk-cache-max-percent added to ADDITIONAL_ARG_VALUE_FLAGS. It takes
  a value, so stripping it from stale Advanced Args without stripping its
  number leaves a bare "10" that argparse binds to the positional model path.
- Added to the command PREVIEW's blocklists and emission, which must mirror the
  launcher exactly or the preview shows flags the launcher strips.

Block-disk store: distinguish "the budget is full" from "this ONE chain is
bigger than the entire budget". Both currently read as the same unactionable
durability-fence warning, but only the second is permanent -- every long prompt
is refused and the cache silently stores nothing. Found by an eviction run with
a 0.5GB budget and a 6050-token/44-layer chain needing ~640MB, where the store
correctly retained nothing and it looked exactly like broken LRU. The new
warning names both numbers and the setting to raise.

TurboQuant contract tests move to the full-precision stored codec for auto mode
on all three families. Measured q4 on vs off: 2.33 vs 2.35s full hit, 2.95 vs
2.97s partial, byte-identical disk -- under 1%, so exactness wins. The LIVE
TQ-KV patch is a separate layer and stays on, which the env assertions still
pin. The explicit-override test is untouched: a user's explicit codec is still
honoured.

Engine 8935 passed / 1 failed (bundled-python drift gate, expected until
re-bundle). Panel 3220 passed.
J
Jinho Jang committed
ae0565fa7da1ead1bab301814cd6f0629d77319f
Parent: 133ac98