fix(shamir): position-based polynomial evaluation; uniform coefficient sampling

[DaBaoAgent]

Fixes #73

Bug

__shamirFn used q.indexOf(a) to derive each coefficient's exponent. indexOf returns the first occurrence of a value, so whenever coefficient values repeat the wrong exponent is used and evaluation is corrupted.

To paper over that, __split_secret forced all coefficients to be distinct from each other and from the secret byte via while (q.includes(w)). This samples coefficients without replacement, violating the uniform i.i.d. sampling Shamir's perfect-secrecy proof requires:

• For t = 3, exactly 3 values per byte are rejected (256 → 253 candidates) → ~0.272 bits of entropy leaked across 16 bytes.
• The leak grows quadratically with threshold t and causes total breakdown at t >= 30.

Fix

1. __shamirFn: evaluate with Horner's method, using the coefficient position as the exponent — identical results for distinct coefficients (verified), correct results for repeated ones.
2. __split_secret: remove the rejection filter; coefficients are now sampled uniformly i.i.d. from GF(256), as required by the security proof.

Verification

• npm test (full suite): 86 passing, 0 failing (includes the existing Secret splitting test, 2000+ iterations).
• New regression tests:
  • position-based evaluation with repeated coefficient values (vs explicit $\sum q_i x^i$),
  • round-trips at thresholds 30–100 (previously broken).
• Backward compatible: shares generated by the old code restore correctly with the new code.
• Issue PoC replicated: old sampling rejects exactly 3 of 256 secrets per byte (t=3); new sampling rejects 0.