fix(observer): report must agree with the predicates at a full window (#735, #736)

[InauguralPhysicist]

#735 — report labelled a still-moving value "equilibrium"

observer_slot_report resolves the six windowed bands in priority order, then falls back to an instantaneous label read off the last dH. That fallback exists for a partial window — where the full-window predicates are false by rule and report still has to say something — but it was never gated on count < N, so it also ran at a full one.

The bands are not exhaustive. A full window of steady gray-band drift at low entropy fires nothing:

predicate	why it's false here
improving / diverging	every step is under dh_small — the #187 gray-band rule
equilibrium / converged	window mean is over dh_zero
stable	requires entropy >= h_low

Every such window was reported equilibrium. A decaying residual is exactly that state:

d is 0.9
i is 0
loop while i < 20:
    d is d * 0.5
    i is i + 1

when (assign count) : 21          <- N = 10, so the window is comprehensively full
report              : equilibrium
equilibrium/converged/stable/improving/diverging/oscillating of d : 0 0 0 0 0 0

So a convergence loop written the way PREDICATES.md recommends — the settled-plus-hold recipe, keyed on report — exits early at rc=0 with a plausible answer. docs/PREDICATES.md:328-335 promised the opposite and named the partial window as the only exception; the observed behaviour is the excluded case. Found by an agent that reimplemented the published formulas as an independent oracle: it confirms the six predicates are right and report is wrong.

Fix. The fallback is gated on a partial window. A full window with no band true reports moving — the residual label the value channel already uses, so both channels answer "not settled, not in any named band" the same way. Relaxing stable's h_low gate was the alternative, and was rejected: the oracle says the predicates are correct, so the wider blast radius would have been spent fixing the wrong thing.

The invariant is now asserted, not described. tests/test_predicate_matrix.eigs checks every full-window shape: report must name a band whose bare predicate is true, or moving when none is. That guarantee was prose only, which is why this survived. Verified red-then-green against the pre-fix build — the assert fires on report=equilibrium with all six false.

Fallout the fix exposed, fixed rather than worked around. lib/experiment.eigs was relying on the wrong answer. All five analyzers seeded tracker is 0 observed, putting a synthetic 0 -> values[0] entropy jump at the head of the trajectory — a real step to every windowed predicate. is_measurement_stable scored ten identical readings as stable only because the fallback overrode the resulting window; with report telling the truth it returned "not stable" for constant data. The sentinel is a declaration, not a measurement, so it is now seeded inside unobserved: and the trajectory is the data.

#736 — four observer papercuts

1. A bare observer query as a statement was silent. W019 already covered the interrogatives (where is x, prev of y); it now covers the query special forms over an ident — report / report_value / observe / trajectory — which print nothing and raise nothing at statement level. That is the worst affordance the observer had: silence is indistinguishable from "the observer had nothing to say", and the bare form is the one issue bodies and READMEs reach for. Zero false positives by construction — over an ident these are compiler-resolved and never reach a user function even when one shadows the name (Name-keyed dispatch superinstruction hijacks a user-rebound 'dispatch' (silent wrong result); also fires on the parenthesized #355 form #459, see W013), and a non-ident argument (report of (x + 0.0)) is an ordinary call the check never sees.
2. state_at leaked __loop_exit__. Now {"x": 4, "i": 0}. The same filter applies to --step's bindings listing, where it also appeared; an explicit p __loop_exit__ still answers, and the tape's binding count stays honest. The filter matches the general __name__ form rather than a hardcoded pair, so a new internal cannot drift back into view — the form is now documented as runtime-reserved in SYNTAX.md, since a user variable shaped that way would be hidden too.
3. SYNTAX.md still called how degenerate — returns 0, 1 only at zero entropy. Observer surface coherence: implement or remove 'how'; decide |x|=1.0; fold in #383 #412 made it a real gradient: 1 - min(1, |dH| / dh_zero), the settledness of the last step.
4. PREDICATES.md still claimed entropy is exactly 0 at |x| ∈ {0, 1}. Observer surface coherence: implement or remove 'how'; decide |x|=1.0; fold in #383 #412 deleted the |x| == 1 special case, so unity is the maximum — the horizon — and only 0 is the home point. Stale in the maximally confusing direction: 0 vs 1 is the difference between "fully converged" and "maximally undecided".

Also adds the missing worked at example inside a loop, where "the last assignment at or before this line" answers with the loop's final pass — precisely where the documented rule stops matching intuition.

Validation

• Release suite 3273/3273.
• ASan/UBSan with detect_leaks=1 3277/3277, leak tally 0.
• New regression coverage: full-window agreement across every matrix shape (test_predicate_matrix.eigs), state_at internal-binding filtering (test_temporal.eigs, queried past EOF so it stays line-insensitive), W019 positives and negatives (test_lint.sh).

docs/COMPARISON.md carries no report-vocabulary claim, so it needed no change; docs/SPEC.md did and got one.

Closes #735
Closes #736

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Fixes observer contract violations and papercuts so report/predicates, temporal queries, and lint/docs all agree with the intended observer model.

Changes:

• Gate observer_slot_report’s instantaneous fallback to partial windows; full windows now return "moving" when no predicate band is true, and add regression coverage asserting full-window agreement.
• Hide runtime-internal __name__ bindings from state_at and unfiltered --step binding dumps via a shared trace_name_is_internal() helper.
• Extend W019 to warn on statement-position observer query special-forms (report/observe/report_value/trajectory over an ident) and refresh observer docs (how, entropy at |x|=1, loop at example, diagnostics/spec/changelog).

Reviewed changes

Copilot reviewed 15 out of 15 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
tests/test_temporal.eigs	Adds regression asserting state_at does not expose __loop_exit__ while still returning user bindings.
tests/test_predicate_matrix.eigs	Adds full-window residual-drift case and asserts the “full-window report agrees with predicates or ‘moving’” invariant.
tests/test_lint.sh	Adds W019 positives for bare observer queries-as-statements and negatives for expression usage.
src/trace.h	Declares trace_name_is_internal() for shared filtering logic.
src/trace.c	Implements trace_name_is_internal() and filters internal bindings from trace_state_at().
src/step.c	Filters internal bindings from unfiltered bindings listing while keeping explicit p __loop_exit__ queries working.
src/lint.c	Extends W019 to warn on discarded observer-query special forms at statement level.
src/eigenscript.c	Fixes observer_slot_report() so full windows with no true predicate return "moving" (no partial-window fallback).
lib/experiment.eigs	Seeds analyzer tracker under unobserved: to avoid synthetic first-step entropy artifacts.
docs/SYNTAX.md	Updates how semantics, adds loop at example, and documents __name__ as runtime-reserved/hidden in dumps.
docs/SPEC.md	Specifies full-window report agreement contract and "moving" residual for non-exhaustive bands.
docs/PREDICATES.md	Documents non-exhaustive full-window “moving” case; clarifies entropy peak at `
docs/DIAGNOSTICS.md	Updates W019 description to include observer query special forms over identifiers.
docs/BUILTINS.md	Updates report vocabulary to include "moving" for the residual full-window case.
CHANGELOG.md	Records both #735 and #736 fixes, including contract clarification and exposed downstream reliance.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 18 out of 18 changed files in this pull request and generated no new comments.