[chaoming0625]

Fixes #184.

Summary

Input-only hidden state (the issue).

• Cause: write_jaxpr_of_hidden_group_transition builds the group jaxpr only from equations reachable from the group invars. A latch latch^t = x @ w that joined the group through V's transition was therefore never produced and never bound, and eval_jaxpr raised KeyError.
• Fix: such an outvar is now bound from the forward pass as a constvar, which gives a zero Jacobian row. This is the mechanism the zero-recurrence fallback already uses.

Pure copies (found while fixing; same family).

• Cause: h2^t = h1^{t-1} (a delay line / shift register) has no equation, so the edge was dropped. h2 fell back to a zero-recurrence group, and the weight feeding h1 got a silently zero online gradient. Downstream's new + 0*old workaround hides this case too.
• Fix: _post_check now merges the copy into the source's group, which gives an identity row.
• A same-step read of the copy is indistinguishable in the jaxpr from a read of h1^{t-1} (they are one variable), so it cannot be fixed there. It now emits a WARNING DiagnosticKind.HIDDEN_COPY_SAME_STEP_READ that names both states. A read on the next step is exact.

Spec: docs/specs/2026-10-24-issue-184-input-only-hidden.md.

Before / after

model (D-RTRL etrace_grad, |grad w|)	main	this branch	BPTT
issue repro (latch)	KeyError	37.508961 (= --workaround result)	—
delay line, copy read next step	KeyError	0.584477	0.584477

Tests

• _compiler/hidden_group_test.py::TestInputOnlyHiddenInMergedGroup:
  • transition + diagonal Jacobian [[0,0],[1,0.9]] for latch <- x@w and latch <- x;
  • two latches;
  • deterministic constvars;
  • shift-register identity row.
• _compiler/hidden_group_test.py::TestHiddenCopySameStepRead: warns for a same-step read, silent for a next-step read.
• _algorithm/d_rtrl_test.py::TestEquationFreeHiddenEdges:
  • D-RTRL equals BPTT through the finite-window oracle (chunked_online_param_gradients, chunk 1 and 3) for the latch and delay-line models;
  • d_rtrl and eprop match the new + 0*old workaround.
• 14 of the 15 new tests fail without the fix.
• Full suite (CPU, remote): 2921 passed, 1 skipped, no failures. The only error is a hypothesis import missing in the remote env, which also occurs on main.

[codecov]

Codecov Report

❌ Patch coverage is 95.00000% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
braintrace/_compiler/hidden_group.py	94.73%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!