fix_moon_phase does not pin ubirthday — wall-clock still leaks into seeded games

[kenforthewin]

Summary

fix_moon_phase=True promises deterministic time-based game effects when seeds are set (per the option docstring: moon phase, friday 13th, night, midnight). The implementation pins the struct tm path via nle_getlt_maybe_fixed() in src/hacklib.c, but ubirthday is still assigned from raw time(NULL) in src/u_init.c, and NetHack hashes ubirthday to vary flavor between games without consuming game RNG:

• src/shknam.c:517 — shopkeeper name selection (nseed = ubirthday / 257)
• src/shk.c:2085 — shop surcharge parity (gameplay-relevant)
• src/mkroom.c:492 — antholemon() ant species (gameplay-relevant)
• src/read.c:72 — scroll label hash
• src/files.c:752 — bones pool selection

getyear() in src/hacklib.c also calls raw getlt(), bypassing the fixed-tm wrapper.

Observed impact

Fully seeded runs (env.seed(core, disp, reseed=False), fix_moon_phase=True, NLE 1.3.0, NetHackChallenge-derived env) of the same seed differ across wall-clock sessions:

• Shopkeeper names change between sessions (e.g. one seed produced "Ms. Possogroenoe" vs "Mr. Kabalebo" vs "Ms. Manlobbi" in sessions hours apart), while turns/score/deaths stay identical.
• A longer/shorter name can push a game message across a display page boundary, changing the keypress sequence an agent needs (extra/missing Space), so byte-exact action-trace replay breaks across sessions. In a 1000-trace corpus we saw a ~1.8% cross-session replay-validation failure rate from this.
• The price-parity and ant-species channels can in principle alter gameplay outcomes, not just cosmetics.

Suggested fix

Derive ubirthday from settings.time_seed when fix_moon_phase && time_seed_is_set, using the same ISAAC64 pattern as nle_fill_fixed_tm(), and route getyear() through nle_getlt_maybe_fixed(). No behavior change when fix_moon_phase is off.

Working patch (verified: same seed run in sessions 96s apart now yields identical action SHA-256 and identical shopkeeper name; no change to any seeded outcome in our regression fixtures): kenforthewin@c8ddc0a

Happy to open a PR from that branch if the approach looks right.