README/docs/index.html: name the specific RLLA mechanism, not just "calibration issue"

Tightens both files' RLLA wording from a general "likely a design/
calibration issue in EpsilonGreedyRLLA's MCS selection" note to the
specific mechanism diagnosed in docs/redesign/P6_reaudit_status.md SS9:
optimistic per-arm initialization (p=0.9) sits almost exactly at the
bandit's own feasibility threshold (0.90), so a single NACK drops an
arm's estimate to 0.855 -- below the cutoff -- blocking greedy
re-selection until a slow, epsilon-driven recovery, systematically
favoring arms far more conservative than the 10% BLER target implies is
optimal. Both files keep the epistemic caveat intact: diagnosed via a
targeted single-seed, two-config trace (SS9), not a full sweep.

Author: karrthikmohank

README.md

@@ -473,9 +473,15 @@ OLLA regulates to 0.99×–1.01× of the 10% target at every config, validating
 it as the baseline. **RLLA**'s realized BLER runs at just **0.13×–0.17×**
 of that target across every config (1.3%–1.7% actual) — a massive
 conservatism, not a reasonable-but-suboptimal tradeoff: it isn't failing to
-decode, it is picking far more robust MCS indices than necessary, likely a
-design/calibration issue in `EpsilonGreedyRLLA`'s MCS selection (which
-specific parameter is unconfirmed). **LTSLA** splits by regime: at 3 km/h
+decode, it is picking far more robust MCS indices than necessary — traced
+(§9) to `EpsilonGreedyRLLA`'s optimistic per-arm initialization (p=0.9)
+sitting almost exactly at its own feasibility threshold (0.90): a single
+NACK drops an arm's estimate to 0.855, below the cutoff, blocking greedy
+re-selection until a slow, epsilon-driven recovery, systematically
+favoring arms with a true success rate comfortably above 90% — far more
+conservative than the 10% BLER target implies is optimal (diagnosed via a
+targeted single-seed, two-config trace, not a full sweep). **LTSLA**
+splits by regime: at 3 km/h
 its BLER matches OLLA's (0.90×–1.02× of target) — a genuine
 spectral-efficiency win at matched reliability — but at 30/120 km/h it
 overshoots to **1.57×–2.07×** of target, worsening monotonically with

docs/index.html

@@ -362,7 +362,7 @@ <h3>Reinforcement-Learning Link Adaptation</h3>
         <span class="metric">LTSLA −1.60% mean / +10.06% best-config vs OLLA (n=5 seeds, N_TTIS=30000)</span>
         <span class="metric">Not clearly distinguishable from zero overall — favorable only at 3 km/h</span>
         <span class="metric">Thompson sampling · ε-greedy · Bayesian SINR posterior</span>
-        <span class="metric">RLLA −37.40% mean — BLER 0.13×–0.17× of the 10% target at every config (OLLA validated at 0.99×–1.01×): massive conservatism, likely an EpsilonGreedyRLLA MCS-selection calibration issue, not a weaker-algorithm story</span>
+        <span class="metric">RLLA −37.40% mean — BLER 0.13×–0.17× of the 10% target at every config (OLLA validated at 0.99×–1.01×): traced (§9) to EpsilonGreedyRLLA's optimistic init (p=0.9) sitting almost exactly at its own 0.90 feasibility threshold — a single NACK drops an arm below the cutoff, blocking greedy re-selection until a slow, epsilon-driven recovery, favoring arms far more conservative than the 10% target implies is optimal (single-seed, two-config check, not a full sweep)</span>
         <span class="metric">LTSLA BLER on-target at 3 km/h (genuine efficiency win) but 1.57×–2.07× of target at 30/120 km/h, worsening monotonically with speed — an under-regulated reliability problem there, not just a suboptimal-but-safe MCS choice</span>
         <span class="metric">Open question, not concluded: the live-Doppler forgetting factor may itself be under-compensating as speed increases — full BLER table & methodology: <a href="https://github.com/karrthikmohank/5g-nr-mac-simulator/blob/main/docs/redesign/P6_reaudit_status.md" target="_blank">status doc §8 ↗</a> (covers only P6's own 6-config × 3-arm grid — no other project's committed data has an equivalent check)</span>
       </div>