A Nat crosses to a JS host as a plain number

[nicolas-abril]

The JS runtime has held Nats as numbers since 87a1e9c (2.3x), and it kept BigInt only at the host seam. This PR removes that seam conversion. An effect's .js gets and answers a number, and a def exported through the .bend loader or -o x.mjs takes and returns one. js_marshal, js_host, nat_host and the marsh memo are gone, which frees 1,163 ttok in comp.ts (63,929 → 62,766).

Breaking for JS callers: a host that passed or expected 5n now uses 5. GUIDE.md, EFFECTS.md, main.ts's header, effs/now.js and the app_win_is_bug_2d demo are updated. A Nat is still capped at 2^48-1 by nat_chk, so no value loses precision as a number.

Tests

• marshal_nat_bigint → marshal_nat_number. marshal_nat_pair and marshal_native_forms now expect number. marshal_array_depth and nat_host_bounds answer numbers.
• Deleted nat_host_neg: a negative Nat answered by host code was caught only by nat_host. JS now runs it raw, the same way the C lane takes the host's word as is.
• Deleted marshal_tag_unknown: the JS lane: a Nat inside a value passed between two loaded .bend libraries stays a BigInt (constructor tags depend on the importing root), since 2.0.29 (still in 2.0.30) #1105 tag check lived inside js_marshal and only ran for types that hold a Nat. The tag problem in JS lane: a Nat inside a value passed between two loaded .bend libraries stays a BigInt (constructor tags depend on the importing root), since 2.0.29 (still in 2.0.30) #1105 itself is unchanged.

This also takes away the Nat half of #1105 and the return-type marshal behind #1150.

Validation

• Generated C: byte-identical to main for all 17 runtime benches.
• Gates on the cluster: repo 54/54, test 1600/1600 (two tests fewer).
• Perf: 3 interleaved gates/perf.ts runs against main. Geomean compiler 0.992, runtime 0.999, no cell over 3%.
• JS export calls: about 4x faster, since there's no BigInt round trip (2M calls: 140 → 36 ms).

🤖 Generated with Claude Code