[InauguralPhysicist]

Integration branch for #1758, Phase 2 numbers: an exact int64 kind beside IEEE 754 doubles, while keeping the observer and deterministic replay. This is a draft; it lands when every round is done and the #1758 Done-when holds.

The plan is docs/design/phase2-numbers.md. Two blind critics reviewed it over three design rounds. Rounds:

• R1, the core model (Alt 1R):
  • R1.1 the slot layer;
  • R1.2 the Value layer (behaviour-preserving);
  • R1.3 the flip: the first int values, tape v7 and observer routing;
  • R1.4 // and floor %;
  • R1.5 bitwise;
  • R1.6 every builtin's int semantics;
  • R1.7 float text and JSON;
  • R1.8 the embed API.
• R2: IEEE doubles, canonical NaN, and writers that refuse non-finite values.
• R3: the observer's distance signal (Observer signal and default: distance (original, O(1)) vs entropy (C-conversion, O(reachable)); opt-in vs always-on #1667), graded by the observer oracle (Observer oracle: one question per predicate, certified labels, per-verdict floor, per-row tier witness #1769).

Landed on this branch so far:

• The design: docs/design/phase2-numbers.md; a 1,798-site inventory checked by tools/phase2_inventory.py; tests/phase2/MAP.md, with tier-witnessed rows per round.
• R1.1, the slot layer.
  • The SMI tag 0xFFFA and its API exist; nothing produces an SMI yet.
  • slot_is_num, SLOT_NUM_RAW and slot_from_num are renamed, and slot_is_numeric and slot_as_double are deleted, so every old reader stopped compiling until given an explicit int decision.
  • New tools: the C slot test (section [0fk]), tools/phase2_check.sh --through <round>, the int-source check, and a corpus differential against main.
  • Local results: suite 4268/4268; differential 0 diffs over 441 programs; calibration 27/27; plant 1 killed.

Related: #1763 and #1774 (float-to-int casts that silently misbehave on large values, which N5 covers); ouroboros#301 (the AOT runtime follows the renames at its EIGS_REF bump).

Closes #1758
Closes #1638

🤖 Generated with Claude Code

https://claude.ai/code/session_01KC99CmwatKssQYggkBCgwF

[codspeed]

Merging this PR will degrade performance by 37.79%

❌ 12 regressed benchmarks
✅ 1 untouched benchmark
🆕 7 new benchmarks

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
❌	bench_idxset	37.3 ms	149.4 ms	-75.05%
❌	bench_dmg_shape	533.7 ms	1,528.4 ms	-65.08%
❌	unobserved armed loop	211.3 ms	427.3 ms	-50.54%
❌	scalar_loop	12.2 ms	19.6 ms	-37.52%
❌	dmg cpu_instrs 2M cycles	1.7 s	2.7 s	-36.83%
❌	bench_dmg_shape (vm-tier)	1 s	1.5 s	-32.8%
❌	dmg cpu_instrs 2M cycles (vm-tier)	1.9 s	2.7 s	-28.69%
❌	bench_perf	99.2 ms	126.5 ms	-21.59%
❌	scalar_loop (vm-tier)	18 ms	21.8 ms	-17.68%
❌	observed_loop	16.4 ms	19.2 ms	-14.15%
❌	unobserved_loop	16.4 ms	19.2 ms	-14.14%
❌	string_build	23.2 ms	26.8 ms	-13.42%
🆕	alloc_heavy	N/A	596 ms	N/A
🆕	bench_buf_mix	N/A	511.3 ms	N/A
🆕	bench_buf_resample	N/A	522.3 ms	N/A
🆕	bench_pcm_codec	N/A	315.6 ms	N/A
🆕	conservative_observed_loop	N/A	29.8 ms	N/A
🆕	conservative_unobserved_loop	N/A	29 ms	N/A
🆕	observer gate soft keyword	N/A	19.2 ms	N/A

Tip

Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing phase2-numbers (5400842) with main (bc50968)}

[InauguralPhysicist]

Open by design: the integration draft for #1758 (Phase 2 numbers). R1.1, the slot layer, is in its critic fix round. Once #1785 lands, this branch takes main and gets a full-lane dispatch (ASan, TSan, macOS). It merges only when every round is done and #1758's Done-when holds.

[InauguralPhysicist]

PRWATCH 2026-11-06: draft, by design. R3.2 (design O3, the four-field O cfg record) was accepted today at 54008422 (#1758): hosted train green except the named CodSpeed exception, blind critic PASS. R3.0-R (#2330, the reveal driver, on main) is with the builder. Remaining before this PR lands: R3.3, R3.4, R5. — Claude DevBox