feat(observer): raw-step signals (#422) + trajectory snapshots across calls (#421)

[InauguralPhysicist]

The pair

Designed together because they interlock: #422 makes the value-channel classifier honest; #421 makes it reachable from the other side of a call boundary. Closes #421, closes #422 — the last two observer-semantics issues from the 2026-07-04 survey.

#422 — the two blind spots, one missing signal

Relative normalization (Δv/(1+|v|)) erased exactly two classes, and both misread as accepting regimes:

Trajectory	Before	After
x → x + 5000 seeded at 4e5 (linear runaway)	stable/converged	diverging
x → -x seeded at 4e-4 (perpetual oscillation)	stable/converged	oscillating
1e6 ↔ 1e6+500 fixed swing	converged	oscillating
x → 2x (geometric runaway)	moving	diverging
damped oscillation / halving / settled constant	converged	converged (unchanged)

The insight: both blind spots are the same missing question. The slot now keeps a parallel window of raw steps and asks whether they are non-vanishing (recent-half mean ≥ half the older-half mean, above a 4·ε·(1+|v|) fp-noise floor). Non-vanishing same-sign steps sum without bound → diverging; non-vanishing alternating steps → oscillating; decaying steps settle exactly as before. The relative step stays the primary contract.

Contracts upgraded to match: expect_converging fails fast on a mid-run diverging, invariant_stable throws on it, and lib/contract.eigs's documented blind-spot caveats are deleted because the instrument no longer has them.

#421 — trajectories as values

Observer state is binding-identity, so a value passed into a function arrives with no history — the naive two-arg contract could never work. Now:

• trajectory of x (special form, two ABI-appended opcodes) snapshots the binding's observer windows into a plain, inspectable dict
• classify of t / classify of [t, "entropy"] (builtin, sandbox-allowed) classifies the snapshot through the same slot machinery report_value/report use — no mirror to drift
• classify of a non-snapshot raises type_mismatch: a bare value never silently classifies as "no trajectory"
• lib/contract.eigs gains the snapshot form expect_regime of [trajectory of x, expected, msg]
• lint knows trajectory as a special form (E003 + the shadow set)

Verification

• Both issue repros pinned red→green in new suite section [50j2] (tests/test_trajectory.eigs, 20 checks), plus no-regression pins for damped/settled/halving, both channels from one snapshot, the loud error paths, and expect_regime
• Release suite 2725/2725; ASan + detect_leaks=1 2725/2725, leak tally 0; both tiers verified (JIT on / EIGS_JIT_OFF)
• SPEC.md + COMPARISON.md gain byte-pinned examples (doc-examples gate green); OBSERVER.md documents the raw-step signal + snapshot model; BUILTINS.md gate green
• test_lint.sh's fixture renamed its classify function — a true-positive W013 now that the name is a registered builtin. Bump-train note: the next consumer sweep should check for user-defined classify/trajectory shadows.
• AOT note: ouroboros pins releases; its aot_rt observer mirror picks this up at the next EIGS_REF bump, per the Observer surface coherence: implement or remove 'how'; decide |x|=1.0; fold in #383 #412 precedent.

🤖 Generated with Claude Code

https://claude.ai/code/session_01GdDAVfZHgoQ2UNP63FYUXm