feat(ml): add IR-drop and thermal ML solver pipeline with validated extraction

[JayRaj21]

Intent

Fix extract_irdrop_batch.sh's liberty/voltage resolution: original script scraped platform config.mk with awk for a plain 'export LIB_FILES = ...' line, silently failing on asap7 (LIB_FILES built via corner/VT-placeholder-substituted make variables) and hardcoded --voltage 1.1 wrong for asap7 (0.77V) and sky130hd (1.8V). Replaced with 'make DESIGN_CONFIG=... print-LIB_FILES print-PWR_NETS_VOLTAGES' (variables.mk's generic print-% target, same mechanism the real flow uses). Also fixed a stdin-draining bug: the make print-X call inherited stdin from the outer while-read loop, and docker run -i reads stdin until EOF, silently draining the loop's process substitution after one iteration; fixed with </dev/null. Verified extract_thermal_batch.sh does not share this bug. Ran full E2E regression (extract_irdrop_batch.sh --force against all 6 real routed designs via real Docker+PDNSim): passed=6 failed=0, every output array byte-for-byte identical to pre-fix baseline. This no-mistakes run's own review step already found and auto-fixed two follow-on issues in commit b6e9947: (1) the make print-X call was a bare assignment under set -euo pipefail that could abort the whole batch on a make failure instead of skipping just that design - now wrapped in if ! ...; then skip; continue; fi like every other per-design step; (2) the make print-X call ran unconditionally per design even when that design would be entirely skipped (outputs already exist, no --force) - now moved inside the irdrop_host skip-check's else branch so it only runs when extraction will actually happen. I manually verified this fix commit: bash -n syntax check passes, and re-ran the batch script without --force against all 6 designs - it now skips instantly with zero docker/make invocations (confirming the idempotent-skip fix), where before this fix it would have spun up one docker container per design just to resolve liberty/voltage even though every design was going to be skipped anyway. The previous run of this same intent failed at the test step not due to a real test failure but because the pipeline's own test agent hit its Claude session usage limit mid-task (after already confirming both review fixes were correctly applied) - this is a re-run of that same validated state.

What Changed

• Added a full ML congestion/IR-drop/thermal pipeline under flow/ml/congestion/ (GNN/U-Net models, training scripts, dataset loaders, inference/prediction, visualization, and an end-to-end run_pipeline.py/run_pipeline.sh orchestrator), plus a flow/ml/Dockerfile and flow/run_visualize.sh for running and visualizing results, and supporting docs (DESIGN_RUNS.md, docs/references/OpenROAD_Thermal_and_Overview_Findings.pdf).
• Added real-data extraction scripts for congestion, IR-drop, and thermal labels/features (extract_labels.py, extract_irdrop_labels.py, extract_thermal_labels.py, extract_netlist_features.py, extract_features.py) and batch runners (batch_run.sh, extract_irdrop_batch.sh, extract_thermal_batch.sh, generate_variants.sh, extract_existing.sh).
• Fixed extract_irdrop_batch.sh's liberty/voltage resolution: replaced awk-based scraping of a plain LIB_FILES line in platform config.mk (which silently failed on asap7's corner/VT-substituted make variables and used a hardcoded, incorrect --voltage 1.1 for all PDKs) with make DESIGN_CONFIG=... print-LIB_FILES print-PWR_NETS_VOLTAGES, the same generic print-% mechanism the real flow uses; also fixed a stdin-draining bug in that call (added </dev/null) that otherwise let docker run -i consume the outer per-design while read loop's stdin after the first iteration.
• Hardened the new make print-X call so a make failure only skips the affected design instead of aborting the whole batch under set -euo pipefail, and moved the call inside the per-design skip-check so it only runs when extraction will actually occur (idempotent --force-less runs no longer invoke Docker/make at all).

Risk Assessment

✅ Low: Commit b6e9947 correctly implements both previously-decided fixes: the make print-X call is now wrapped in an if-guard (set -e safe) that skips just the offending design instead of aborting the batch, and it is relocated inside the irdrop_host skip-check's else branch so it only runs when IR-drop extraction will actually happen; no new logic gaps, scope issues, or regressions were found, and the change matches the user's stated intent with no contradictions.

Testing

Ran the actual review-fix commit (b6e9947) end-to-end with real Docker/OpenROAD against a synthetic bad-config design and a valid nangate45/gcd design: confirmed a make print-X failure is now contained (skip + continue, batch completes normally) rather than aborting the whole script, and confirmed an idempotent rerun with outputs present triggers zero docker invocations, both matching the two auto-fix decisions from review round 1. All test artifacts (fake results/, designs/, and ml/congestion/data/ files) were removed afterward; working tree is clean.

Evidence: Real docker run — bad config.mk contained, batch continues (no set -e abort)

Design: nangate45__badcfg_base
[SKIP] features already extracted
[SKIP] 'make print-LIB_FILES print-PWR_NETS_VOLTAGES' failed for designs/nangate45/_badcfg/config.mk
======
Design: nangate45_gcd_base
[SKIP] features already extracted
[RUN] extract_irdrop_labels.py ...
(real make print-X resolved NangateOpenCellLibrary_typical.lib, voltage 1.1V; extractor itself failed only because I used empty placeholder odb/spef files, not real routed data)
======
Done. passed=0 failed=1 skipped=3

Evidence: Idempotent rerun without --force — zero docker invocations for make print-X

Design: nangate45__badcfg_base
[SKIP] features already extracted
[SKIP] IR-drop labels already extracted
======
Design: nangate45_gcd_base
[SKIP] features already extracted
[SKIP] IR-drop labels already extracted

Done. passed=0 failed=0 skipped=4
(docker ps -a diff before/after: no new containers)

Pipeline

Updates from git push no-mistakes

✅ **intent** - passed

✅ No issues found.

✅ **Rebase** - passed

✅ No issues found.

✅ **Review** - passed

✅ No issues found.

✅ **Test** - passed

✅ No issues found.

• bash -n ml/congestion/data_collection/extract_irdrop_batch.sh (syntax check)
• Real docker+OpenROAD run of extract_irdrop_batch.sh against a design with a deliberately broken config.mk ($(error ...)) plus a valid nangate45/gcd config, confirming the bad design is skipped ([SKIP] ... failed) and the batch continues to and correctly resolves liberty/voltage for the next design, ending with a normal 'Done.' summary rather than aborting
• Real docker+OpenROAD rerun of the same batch without --force with all outputs already present (docker ps -a diffed before/after), confirming zero docker containers are spun up for make print-X and every design instantly skips

✅ **Document** - passed

✅ No issues found.

✅ **Lint** - passed

✅ No issues found.

✅ **Push** - passed

✅ No issues found.