RFC 0002 wave 1b: surprise, pace, dispositions (0.10.0)

[BitmapAsset]

Wave 1b of RFC 0002 (docs/rfcs/0002-neuromodulators-and-body.md, Designs B, C and D), stacked on #39 (wave 1a): the surprise, pace and dispositions blocks.

1. PR RFC 0002 wave 1b: surprise, pace, dispositions (0.10.0) #40 (draft) this PR, head 509c8d04260c8563652cfbf655712b668cc73342 on rfc2-wave1b-surprise-pace-dispositions, base rfc2-wave1a-needs @ 8e2600c (stacked; rfc2-wave0-freeze 889a993, main 164bd3b); CI on the full head sha: ci run 38887899248 success, check run 38887899289 success. Never marked ready, merged, tagged, released or posted.
2. Tests: new unit, live, monitor and golden tests all pass (9 unit tests covering every closed form of Designs B-D, the α⁻ seam, 30 refusal cases for V-R1, V-R2, P-R1-P-R3, D-R1, D-R2, 18 state refusals, surfaces, prove boxes, the refusal property); the whole suite cargo +stable test --workspace --locked --no-fail-fast locally: 32 targets, 404 passed, 3 failed, 1 ignored, the 3 being the wall-clock tests (acc1_tv_le_001_within_10ms, sample_on_budget_is_a_deadline_at_many_chains, decide_fallback_enumerates_a_program_no_chain_can_start) that fail on the base's release build at the same moments too and pass in CI; MSRV 1.78 exit 0 with Cargo.lock unchanged; P0 653 outputs, 0 differ, 8/8 strands.
3. Bench (bars v2, the frozen release binary of the head, PROBBIT_THREADS=1; nulls.py 51 rows, 0 pass): s1 pass (ratios 0.5-0.6667); s2 pass (1.0343-1.0719); s3 pass (1800/1800, error 0.0); s4 pass (1800 strands 0 diff); s5 pass (P0); p1a pass (error 0.0, 0 changed); p1b pass (0 changes, control 0.0); p1c pass (Spearman 0.8414); p6 pass (transients within 1.1e-5, pins, burst exact); o1 pass (pooled -0.001417, gaps 0.0); o2 pass (100/100, 0.0); o3 pass (distance 0.002753, 18/20); o4 pass (0 breaks); o5 pass (4/4); wave 1a's bars re-measured on this build: n1, n2, n3, n4, n5, n6, n7, n8, n9, v1, v2, v3, om1 pass; n10, n11 fail as in wave 1a (the design findings it recorded; bars unchanged).
4. Changes after the freeze: scripts 0, fixtures 0, MANIFEST unchanged, bars.md / bars.json / bars-v1.md byte-identical; RFC text: item 5 of the dated note (P-R3 also refuses a habit condition on pace_up / pace_down, and an event that sends either is refused at inputs.<id>; the stance's surprise map is {} once every gate is closed; precision reads each sense's variance after the event's sense step; the pace clause prints R̄ to 2 decimals). No pass value moved.
5. Risk: a new learner's gate starts open: g = (r_ref + 1) / 2 = 2.5 at the defaults, so every new individual takes steps up to 2.5x the fixed rate's on its earliest feedback and its why says learning fast: feedback changed before any feedback has changed; the bars measure what the gate buys on a reversal (s1) and what it costs on a stable phase (s2: 1.03-1.07x), not how a host reads the clause on a brand-new individual. Wave 2 needs: load_level is an engine-owned number input like pace_up / pace_down: reuse build_body's lowering (reserved only with the block, refused when sent, left out of inputs, boxed by prove); the load step goes right after pace_step in prep; a rest line must say what it does to each learned trait's gate (v, r̂, g, counts sit next to Δ and the RFC is silent); load.from.expired will read credit expiries, which the patience gene now scales.

Bench (bench/body, bars v2, the release build of the head)

python3 bench/body/nulls.py: 51 competent nulls over 40 bars, 0 pass (exit 0). Then the parts of run.py --only needs,gate,pace,dispositions,valence,replay,habits,p0, with PROBBIT_THREADS=1 (outputs are byte-identical at any thread count); the gate part's nine (T, noise) cells ran in parallel through the part's own functions and were combined by its own gate_cells.

bar	measured	bar value	result
s1 flip re-learning	gated / fixed median turns per (T, noise): T100 n0.00 0.625, T100 n0.10 0.5, T100 n0.20 0.5769, T40 n0.00 0.625, T40 n0.10 0.5556, T40 n0.20 0.6667, T60 n0.00 0.625, T60 n0.10 0.5, T60 n0.20 0.625; every same-budget constant gain re-learns strictly slower (gated50 medians 5.0, 5.0, 7.0, 5.0, 6.0, 8.0, 5.0, 5.0, 7.0)	≤ 0.75 in every cell; constant gains within the gate's movement strictly slower	pass
s2 stable-phase movement	gated / fixed phase-A movement: T100 n0.00 1.0502, T100 n0.10 1.0343, T100 n0.20 1.0719, T40 n0.00 1.0502, T40 n0.10 1.0415, T40 n0.20 1.0613, T60 n0.00 1.0502, T60 n0.10 1.0383, T60 n0.20 1.0626; mid on top at the end of phase A in 1.000, 1.000, 1.000, 1.000, 1.000, 1.000, 1.000, 1.000, 1.000; re-learned before T in 1.000, 1.000, 1.000, 1.000, 1.000, 1.000, 1.000, 1.000, 1.000	≤ 1.5; ≥ 0.9 and ≥ 0.9	pass
s3 caps and the opening step	1800/1800 lives within g_max × step_cap and ±total_cap; opening step max error 0.0	100 %; within 1e-5	pass
p1a pace reads the past	max	R̄ − recomputed	0.0 over 2,000 events; 0 earlier stances changed by a perturbed future (4 cuts)
p1b offered stakes do nothing	0 pace changes over 10 fuzzed streams; the accepted-win control moves R̄ by 1 − 0.5^(1/6) within 0.0	0; within 2e-6	pass
p1c pace tracks the reward rate	Spearman 0.8414 on blocks; bursts: mean steady 0.499996, bursty 0.5, bursty range 0.401242, spaced last 0.558741 vs 0.558743	≥ 0.7; within 0.02; ≤ 1.0; within 1e-4	pass
p6 loop gain	G 0.24: 1.263152 vs 1.263158; G 0.6: 1.499989 vs 1.5; G 0.84: 2.249084 vs 2.249084; pinned G 1.08 a_cap 12: 3.599996 (max 3.599996); G 1.08 a_cap 20: 3.999996 (max 3.999996); G 1.2 a_cap 12: 3.599996 (max 3.599996); G 1.2 a_cap 20: 3.999996 (max 3.999996); burst: max 4, 4 from the 10th win, recovery error 0.0	within 1e-4; at the binding cap, never above 4; exact recovery within 2e-6	pass
o1 E settles at the expectile	pooled sign-aligned error -0.001417 over individuals 0-4 (o: -0.2877, 0.2429, -0.1708, 0.1468, 0.2556); gaps stream max error 0.0		·
o2 the gene's sign is readable	100/100 individuals; trajectory max error 0.0	100/100; within 1e-5	pass
o3 twin distance rises with outcomes	mean distance 0.002753 at spread 0.3 (0.0 at spread 0); 0.0 up to the earliest win; pursue differs in 18/20 pairs	≥ 0.001; 0 before; ≥ 10 of 20	pass
n5 0 habit breaks	{"n5": {"bound": 12600, "breaks": 0, "habit_free_twin_breaks": 1317}}	0; ≥ 50 bound	pass
o4 0 habit breaks	{"o4": {"bound": 2390, "breaks": 0, "habit_free_twin_breaks": 1709}}	0; ≥ 50 bound	pass
n4 replay	112 strands: 0 verify failures, 0 third-copy and 0 split-copy differences; planted changes refused: {'changed': True, 'removed': True, 'swapped': True}	0 differing bytes; nulls refused	pass
s4 replay	1800 strands: 0 verify failures, 0 third-copy and 0 split-copy differences; planted changes refused: {'changed': True, 'removed': True, 'swapped': True}	0 differing bytes; nulls refused	pass
n6 / s5 P0	corpus exit 0: {"outputs": 653, "differ": 0, "identical": 629, "stderr_warnings_only": 24, "strands_verified": 8, "verify_failed": [], "seconds": 1753.2}	byte-identical	pass (n6 pass)
o5 P0	4/4 pairs identical without digests; a planted change detected: True	spread 0 = today's bytes	pass

Measured tables: BENCHMARKS.md §12 at the head.

Changes after the freeze

• Scripts, fixtures and MANIFEST: none; bars.md, bars.json and bars-v1.md byte-identical to the base.
• RFC text: item 5 of the dated note (P-R3 also refuses a habit condition on pace_up or pace_down, and an event that sends either; the stance's surprise map is {} once every gate is closed; precision reads each sense's variance after the event's sense step; the pace clause prints R̄ to 2 decimals). No pass value moved.

Draft: never marked ready, never merged.