t27b: resolve cross-module use names - the 28 ml specs #8993 unlocked on the reference #9065

What

Master's #8993 taught the t27c test-report (the t27b corpus reference) to resolve cross-module use imports. 28 specs/ml/** specs now PASS the reference and BLOCK on t27b, which still reports the imported names as undeclared:

• 26 x ExprCall(undeclared fn)
• 2 x ExprIdentifier(undeclared)
• 1 x ExprArrayLiteral(repeat count)

Measured on the t27b lab (branch t27b/stdmem-parity-final, master cc19d1f + #9056, clean corpus sweep, fresh reference cache):

reference_pass 1426   t27b_pass 1398   blocked-on-t27b 28 (all specs/ml/**)   crash 0   per-test disagree 0

The 28

activation: elu, gelu, gelu_approx, silu_swish, softmax, tanh; igla_champion_capsule; layers: batchnorm, embedding; loss: binary_crossentropy, contrastive, cross_entropy; optimizer: adagrad, adam, lamb, race_config; pathway/mlp; recurrent: attention_mechanism, bilstm, gru_cell, lstm_cell, lstm_single, rnn_cell; rl: advantage_estimator, ppo_clip_loss, ppo_critic; transformer: encoder_block, positional_encoding.

Scope

Teach t27b's frontend the same cross-module name resolution the reference now has (the constructs live in the specs' own use graphs), then re-sweep: the gate is t27b_pass == reference_pass on the full corpus with 0 per-test disagreements. The parity ledger and sweep harness are in place from #6063.

🤖 Generated with Claude Code

User Scenarios

• Given the 28 specs in specs/ml/** that use cross-module use imports, when t27b's frontend resolves those imported names the same way the reference does, then t27b no longer reports ExprCall(undeclared fn), ExprIdentifier(undeclared), or ExprArrayLiteral(repeat count) errors for those specs.

• Given a full corpus sweep with the parity ledger from EPIC: t27b runs the whole corpus -- 100% of reference-passing specs, 0 mismatches, experiments on Railway #6063, when t27b is updated with cross-module name resolution, then t27b_pass == reference_pass with zero per-test disagreements.

Requirements

• FR-001: t27b's frontend MUST resolve cross-module use imports by traversing the use graphs defined in the specs themselves, matching the reference's behavior added in specs/ml tests run: 40 more specs green (10 -> 50 of 61); use resolution and two Zig emitter fixes (refs #8980) #8993.

• FR-002: The resolution MUST handle the three error classes currently blocking the 28 ml specs: ExprCall(undeclared fn) (26 occurrences), ExprIdentifier(undeclared) (2 occurrences), and ExprArrayLiteral(repeat count) (1 occurrence).

• FR-003: The implementation MUST NOT change the reference or the 28 ml specs; it MUST only modify t27b's frontend name resolution.

• FR-004: After the change, a clean corpus sweep MUST produce t27b_pass == reference_pass with per-test disagree 0 and crash 0.

Success Criteria

• Running the corpus sweep reports t27b_pass 1426 (matching reference_pass 1426) with blocked-on-t27b 0 and per-test disagree 0 — exit 0
• The 28 specs under specs/ml/** individually pass t27b without undeclared-name errors — 28 tests
• scripts/tri_loop/t27b.py compiles and runs the sweep without crashes — exit 0

Boundary

• scripts/tri_loop/t27b.py
• experiments/gfternary-line/resolve.py