Recompute tuning consensus before validation

[thomasahle]

Follow-up to #261.

The frozen combiner already applies the preregistered support/L1/SHA rule, but the independent prevalidation audit previously trusted its selected field. This adds a standalone audit that canonicalizes all 20 lane optima and recomputes both the winner and its evidence before held-out games can be read.

It also reports the compact candidate, disabled mechanisms, maximum real-edge cost, and finite null/cap horizons needed by the formal port.

Verification:

• combined tuner suite: 161 passed
• protocol discovery suite: 10 passed
• synthetic 10:5:5 consensus accepted
• tampered selected vector rejected
• git diff --check clean

No engine code or tuning result changes.

[thomasahle]

Follow-up audit found one more pre-result edge case: if the literal consensus selected IID=1, the compact auditor preserved it even though the preregistered source-cost rule says to drop IID when enabling it adds engine lines. Commit 5e67ebc now applies that rule explicitly, resets inactive IID subparameters, and reports literal and compact mechanism states separately.

Validation: 163 tuning tests plus the complete repository suite (580 passed, 2 skipped). No game observations or selection rules changed.