feat(sim): ordered analysis_steps[] -- several named solves per corner

[loom-fleet-dispatch]

Closes #2482

Adds an additive, ngspice-only analysis_steps[] request field: ordered, named solves per corner with per-step source overrides, step-scoped measurements, and derived cross-step measurements graded through the existing limits machinery. Scalar analysis is unchanged. Validation (duplicate/unknown/forward refs, bad overrides, ambiguous names) happens before dispatch; a failed step marks dependent derived measurements unavailable; Xyce is refused clearly; steps are part of the checkpoint fingerprint; remote/batch work unchanged (deck generated worker-side).

Files: src/klayout_tools/sim_steps.py (new), sim.py, sim_remote.py, op_sanity.py, docs/cli/sim.md, CHANGELOG.md, tests (tests/test_sim_steps.py, test_sim.py, test_sim_batch.py, test_op_sanity.py).

Tests: tests/test_sim*.py tests/test_op_sanity.py all pass.

🤖 Generated with Claude Code

[loom-fleet-dispatch]

Judge review: APPROVED (head 98a723a)

Verified against each #2482 acceptance criterion in a judge worktree:

• Scalar compat: analysis path unchanged (_corner_measurement_results original branch; report gains no step fields -- test_scalar_run_report_gains_no_step_fields). analysis + analysis_steps together is refused.
• Ordered named steps: new analysis_steps[] with unique names, per-step alter, step-scoped expr/.meas, derived top-level measurements[]; one .control block, each step in its own fresh plot with klt_plot_<i> capture (guards against mispairing with prior step's solve).
• Pre-dispatch validation: duplicate step/measurement names, unknown/forward refs, invalid alter targets (also checked against the netlist and corner supply axis), ambiguous log keys, empty sequence, meas-type/kind mismatch -- all raised in resolve_analysis_steps/validate_alter_targets before dispatch.
• Provenance and failure propagation: step/derived_from on measurements and rollups; steps[] per corner; step_failed vs derived_input_unavailable vs ordinary measurement diagnostics; dependents never graded; real-ngspice tests for failed intermediate step and failed derived expr.
• Checkpoint: complete ordered definition is in analysis hashed by _checkpoint_fingerprint; test covers alter/args/order changes; resume test present.
• remote/batch: request forwarded verbatim; test_run_sim_batch_accepts_an_analysis_steps_request and remote doc test.
• Xyce: refused up front, checked first. options.waveforms/--plot are also refused for steps (not silently ignored).
• Docs: docs/cli/sim.md has the new section, request/response field tables, diagnostic codes; CHANGELOG updated. docs/json-contract.md untouched, but sim fields are documented in sim.md and changes are additive (non-blocking).
• Tests: test_sim_steps, test_sim, test_sim_batch, test_op_sanity, test_sim_staging, test_sim_save_mode all pass locally (no failures); ruff check src tests passes cleanly (no new UP038 noise). CI: Lint, golden, docs/contract checks green; Python 3.10-3.14 test matrix was still in progress when reviewed, so mergeStateStatus is BLOCKED pending those -- confirm green before merging.

[loom-fleet-dispatch]

Automated re-date of stale required checks (#8508), 1 of 3

The #8248 required-check-freshness guard blocked the merge of PR #2988: a required check's green result predates the current base-branch tip, and this repo's merge token has no actions:write to re-run that check directly. Pushed a tree-identical no-op commit (96fb3c8) instead — the diff is byte-for-byte unchanged, but every required check now re-runs against a current timestamp.

This moves the head SHA, which invalidates the standing Judge approval (#5686). Expect this PR to cycle back through loom:review-requested once CI on 96fb3c8 completes, then merge normally once re-approved.

The #8248 guard itself is unchanged and still applies to the next merge attempt. This is re-date 1 of a budget of 3 for this chain of tree-identical heads (#9590); once the budget is spent, a further block escalates to loom:operator.

loom dashboard

[loom-fleet-dispatch]

Chain-head merge lock (#10167): this PR was just re-dated at 96fb3c8. Other PRs targeting main defer their merge (merge-pr.sh exit 6) until it lands or closes, its head moves, a required check fails, or 1200 s pass, whichever is first.

loom dashboard