feat: add seeded Thompson-sampling router policy (#48)

[Nitjsefnie]

Closes #48.

Adds a thompson routing policy beside epsilon-greedy (default) and ucb, mirroring the structure of the UCB router from #44 — same POLICIES registry, same per-citizen sample-count aggregation across required_tags, and --explain naming the active policy.

Formulation (Gaussian Thompson sampling): each citizen's combined score is the posterior mean; the width is THOMPSON_EXPLORATION_C / sqrt(n) (c = sqrt(2), matching UCB_EXPLORATION_C), with unplayed citizens (n=0) getting a wider fixed prior (sigma=2.0) so posterior width shrinks monotonically with evidence and cold-start explores everyone broadly. Sample one value per citizen, pick the max, ties break by citizen id.

Determinism (the crux — sampling is stochastic, and the suite must be reproducible): pick_thompson/thompson_sample_values take an injected random.Random — never global random. recommend(seed=) and main() build it from --seed, the exact convention epsilon-greedy already uses. Citizens are sampled in citizen-id order, so the draw sequence (and every sample) is a pure function of the seed, independent of dict order — a given (state, seed) is fully reproducible. The flag-free epsilon-greedy default path is textually untouched and byte-identical across fixed seeds (0/1/7/42/999).

Tests (ThompsonRoutingTest in scripts/test_lessons_routing.py, mirroring UCBRoutingTest): monotonic posterior-width shrink, 20-call seeded determinism, order-independence of the sampling seam, cold-start spread across unplayed arms, exploit-the-leader on ample evidence, empty-input safety, and a decision test pinned to the mean/evidence accounting. That last one bites: dropping the mean from the sample (rng.gauss(score, …) → rng.gauss(0.0, …)) flips the chosen arm and fails the test; restored, all 21 pass. references/routing.md gains a Thompson subsection beside UCB and the README policy list is updated.

All CI test scripts pass on both main and this branch, and --policy thompson --seed 7 --explain produces a reproducible pick with the sorted sampled-value breakdown.

Disclosure: this contribution was made with AI assistance (Claude), verified against the repo's full test suite and mutation-tested before submission.

[vercel]

@Nitjsefnie is attempting to deploy a commit to the yehudamoshe24-6977's projects Team on Vercel.

A member of the Team first needs to authorize it.

[ymxlx]

Verified locally against the research-team example: 21 routing tests pass (7 new Thompson cases), --policy thompson --seed 7 is reproducible across runs, and the flag-free epsilon-greedy default path is byte-identical. Clean mirror of the UCB structure (#44) — injected RNG, citizen-id-ordered sampling for seed-determinism, mutation-pinned decision test, docs + roadmap updated. Thanks for the thorough writeup and the disclosure. Approving.