docs+: P3 transparency cleanup (round-bankers, rd.log(-inf), stale docs)

[sigilante]

P3 — docs/traceability batch (Gnome/Angel review)

Mostly comments/docstrings/READMEs, plus two trivial correctness one-liners with tests.

Behavioral (with tests) — math.hoon

• round-bankers now actually rounds half-to-even. It was half-up: it took the already-toi-rounded value and bumped ties up, so round-bankers(2.5) → 3 (not 2). toi under %n ties to even, so force ^rs/^rd %n. → test-round-bankers (2.5→2, 1.5→2, 3.5→4).
• rd.log(-inf) returns NaN (was 0.0 — a copy-paste from exp; rs.log was already correct). Also fixed the stray exp(...) comments in the rd.log branch. → test-log-ninf.

Docs / comments

• math.hoon tiny (rs/rh): docstring/example said "smallest normal" but the value is the smallest subnormal (2^-149 / 2^-24) — corrected, with a pointer to the normal.
• sur/lagoon.hoon: %cplx listed @cs/@cd → now @ch/@cs/@cd/@cq; augmented the tail=* comment to explain it's deliberate per-kind specialization space (slots in without rewriting jet axes — per the design rationale; %fixp uses it for precision).
• lagoon.hoon: documented +check (the 0x1-pin / prod(shape) invariant every public arm relies on) and the known %i754→%uint negative-value limitation (was a bare XXX).
• saloon.hoon: +neq doccord said "equal to" → "not equal to"; fixed the data=data= typos in the gth/gte/geq examples.
• READMEs: lagoon Type System refreshed (%real→%i754; all six kinds shipped, not "planned"); saloon gains a linear-algebra (eig/eigvals/eigvecs) section.

All edits are in regions disjoint from the open PRs (#50–#53), so no conflicts expected. math.hoon, lagoon.hoon, saloon.hoon all rebuild green; the new math-rounding tests pass.

(Deferred to the #18 Chebyshev rewrite, already noted there: the log hang, exp cancellation, asin/acos & sqt(-0) crashes, and the no-op NaN guards.)

🤖 Generated with Claude Code