[oharboe]

Removes bazel-orfs, bazel-orfs-verilog and orfs (plus yosys, yosys-slang, rules_verilator and verilator, which only existed for them) from MODULE.bazel, and deletes test/orfs/ (gcd and mock-array), along with the two verilator dev-dep checks in test/BUILD.

Side effect on version selection: rules_bison/rules_flex 0.3.1 and rules_m4 0.2.x do not load under Bazel 9 (CcInfo removed, missing //bin:m4). yosys had been masking this by raising them through MVS. This PR declares the versions master actually resolved: rules_bison/rules_flex 0.4 and rules_m4 0.3. libffi resolves to 3.4.7.bcr.3 instead of bcr.4.

Testing strategy
From now on, OpenROAD relies on fast, single-concern tests for its promises, not on flow-level integration tests. Each test asserts one intended behavior on small inputs and runs in seconds. Writing such tests used to be the expensive part; with AI-assisted development it is now cheap, so we write them when we need them rather than keep a heavyweight integration suite that approximates them. Full-flow integration testing stays in ORFS, where the flow lives. OpenROAD CI stays fast and lightweight.

What prompted this
Regenerating rules-base.json used to be a local-only flow: bazelisk run //test/orfs/gcd:gcd_update, then commit the file. That is no longer possible.

• ORFS [Docs] Fix OR messages glossary #4552 (76be2a8eb, merged 2026-09-23) removed the rules files, checkMetadata.py, genRuleFile.py and RULES_JSON. It replaced them with make metadata-check, which posts metadata.json to the QoR dashboard server and compares it with the dashboard's latest master build.
• That check needs the network during the test, gives a different answer as the dashboard moves, and cannot compare a change with its merge base. bazel-orfs therefore dropped its rules-file test and update targets instead of adopting it (bazel-orfs 9347dbeb).
• At our current pin, gcd_update already fails locally (cp: target '.../rules-base.json': Not a directory). bazel-orfs fixed that in bad2c699, after our pin 551d4c72, and picking up the fix means bumping into the ORFS change above.

Keeping test/orfs would mean either staying pinned to a stale ORFS/bazel-orfs indefinitely, or taking a server-dependent QoR gate into OpenROAD CI. The first leaves test/orfs going stale. The second adds more CI brittleness to an already fragile multi-repo chain. Neither is worth it.

About mock-array
mock-array has traditionally been a highly valuable regression test. Its regular N×N array of identical, abutted macros built from a hierarchical block is simple RTL, but it hits corners that production SoCs hit. It exposed:

• MPL fixes: d1dd49a, 52c6ce2, f7cbe0a, 2ae9067, 38bf6ed, mpl: shrink pin access area if design does not fit #8366
• CTS skew and hold problems in macro-dominated designs: cts: Large clock skew in asap7/mock-array design #8255, Hold problems in CTS #8516, Improve CTS so that default CTS_ARGS gives as good results as tweaked CTS_ARGS in mock-array #4989
• hierarchical-flow bugs: read_db -hier problem with mock-array power test case #8141, reg2reg path group is missing with read_db -hier #8166, floorplan: No TAPCELL insertion in asap7/mock-array design w/ the hierarchical flow #8254, cts: Incr STA vs. full STA timing difference after CTS #8240
• netlist/SPEF name mismatches: -hier spef and .v file don't match up #8297, Many warning due to net name mismatch b/w netlist and SPEF #8551
• sta: read_vcd crash in bazel test w/ test/orfs/mock-array/power.tcl #8857, Verilator failed with 17k warnings in asap7/mock-array hierarchical flow #8108, and write_macro_placement escaping (475eb39)

It is, however, a poor fit for the ORFS Makefile flow. It needs bazel-orfs for:

• a separate synthesis run per module, with netlists stitched together
• a block → abstract → parent dependency
• parameterized 4x4 and 8x8 variants
• post-flow orfs_run checks

That is why it lived in OpenROAD, and why OpenROAD had to carry bazel-orfs.

If we find we need mock-array coverage again, we will resurrect mock-array in bazel-orfs, which already builds OpenROAD from source and owns the flow interface. We will then have AI generate specific single-concern OpenROAD test cases for whatever it catches.

What we get immediately

• No more bazel-orfs/ORFS bumps in OpenROAD. There were 14 since 2025-09, and about 200 commits touched test/orfs/, many of them just regenerating rules-*.json after yosys and STA bumps.
• No OpenROAD PR is needed when the ORFS interface changes, so Bazel-breaking-changes.md goes.
• A smaller dependency graph: no git fetch of bazel-orfs during module resolution, and no yosys, slang or verilator.
• bazelisk test ... gets faster; the mock-array 8x8 flow was the slowest target.
• No dependency on the network-backed QoR dashboard check that ORFS [Docs] Fix OR messages glossary #4552 introduced.

Lost coverage: the risk we accept
Flow-level breakage in general is still caught by ORFS's daily OpenROAD-bump PR (github-actions-cron-update-OR.yml), within a day but after merge. The following is exercised by test/orfs and by nothing else in OpenROAD tests or public ORFS CI. It is lost until mined back.

Integration level

Lost	Was exercised by
Pre-merge full-flow smoke test inside bazelisk test ...	gcd, mock-array
End-to-end hierarchical macro-array flow: hierarchical block P&R → abstract → N×N abutted array → P&R → post-P&R STA handoff. No ORFS design has this shape	MockArray_{4x4,8x8}_base_test
The same macro-array netlist in both hierarchical (OPENROAD_HIERARCHICAL=1) and flat mode	MockArray_*_{base,flat}_test
CTS and hold repair through narrow channels into an array of macro clock pins, inside a real flow (cts/test/array* covers CTS alone)	MockArray flow

Specific concerns

Lost	Was exercised by
Determinism: identical runs give byte-identical .odb/.sdc at floorplan, place, cts, grt, route, final	gcd_idempotent_test_* (check_same.*)
whittle.py end-to-end with the real openroad (etc/test_whittle.py only uses mocks)	gcd_whittle_test
OpenROAD-written hierarchical Verilog + SPEF + SDC reload in a fresh OpenSTA session, read_spef -path per instance, no unexpected warnings (#8297, #8551). The existing tolerance of pin * not found (io_lsbOuts_0) needs a look	MockArray_*_path_groups_test, load_mock_array.tcl, parasitics.tcl
find_timing_paths -from/-to and group_path -through on those reloaded outputs	path_groups.tcl
place_macro with escaped hierarchical names (ces_row\[r\].ces_col\[c\].ces), exact location and orientation, grid invariants (count, one orientation, aligned rows and columns, X abutment)	place_mock_array.tcl, check_macro_layout.tcl
write_macro_placement with escaped hierarchical names	write_macros.tcl
-mirrored_pins spread over several layers per direction	element-io.tcl
Abstract LEF of a hierarchical (OPENROAD_HIERARCHICAL=1) block used as a macro	Element_*_generate_abstract
Connectivity by abutment of mirrored macro pins (zero-length nets, DRC = 0)	MockArray flow
PDN over an abutted macro array with a tight channel ("PDN problems if it is smaller. Not investigated.")	BLOCKS_grid_strategy.tcl, array_spacing_*

Accepted without replacement

Lost	Was exercised by
QoR thresholds and netlist hashes in rules-*.json (they only record current output, and ORFS #4552 has retired rules files upstream)	gcd, gcd_b, 4 MockArray variants
8x8 scale	MockArray_8x8_*
DETAILED_ROUTE_END_ITERATION	mock-array
gcd with OPENROAD_HIERARCHICAL=1, explicit die area, ABC_CLOCK_PERIOD_IN_PS and an env-driven SDC (ORFS/yosys features; ABC_CLOCK_PERIOD_IN_PS is only used by gf12)	gcd_test
Slang frontend + SYNTH_BLACKBOXES + stitched SYNTH_NETLIST_FILES in one design (each is covered separately by ORFS asap7/uart and asap7/aes)	mock-array synthesis

Everything else test/orfs touched is covered by existing OpenROAD tests or ORFS designs:

• OpenROAD tests: ppl/test/write_pin_placement*, ppl/test/annealing*, tap/test/abutting_macros_step*, tap/test/cut_rows_halo, cts/test/array*, gpl/test at density 0.8, dbSta/test/read_vcd, OpenSTA search_timing_model_readback
• ORFS designs: asap7/gcd, asap7/swerv_wrapper, asap7/aes-block, asap7/riscv32i-mock-sram, asap7/uart, asap7/aes, gf180/aes-hybrid

Not lost, because it never actually ran:

• VCD power (power*.tcl) and Verilator netlist checks were skipped (needs_vcd → continue, ELEMENT_POWER_TESTS = []).
• RTLMP was bypassed by MACRO_PLACEMENT_TCL, so the MPL fixes above are already unguarded.
• path_groups ran under bare OpenSTA, not openroad, so reg2reg path group is missing with read_db -hier #8166 was not really covered.

Recovering coverage: mine on breakage
test/orfs stays in git history at this PR's parent commit (6e51da9). When master breaks on one of the concerns above (in the ORFS bump PR or a user report), the fix PR adds a single-concern regression test mined from that history with an AI coding assistant. If mock-array as a whole turns out to be needed, it is resurrected in bazel-orfs (see above). Suggested prompt:

Master broke on <concern / ORFS PR link>. Mine test/orfs at 6e51da991dcb
(git show 6e51da991dcb:test/orfs/...) for the matching case and write a
single-concern OpenROAD regression test: seconds to run, no yosys/ORFS,
checked-in or Tcl-generated inputs, asserts intended behavior
(check_log = False, check_passfail = True, or gtest; no new .ok/.defok
golden), registered in CMake and Bazel. Show it failing on the broken
commit and passing with the fix.

Verification

• bazelisk build //:openroad and bazelisk test ... --keep_going pass locally (4437 pass; the 2 failures were the verilator dev-dep checks, removed in the second commit and re-verified with //test/...).
• bazelisk mod graph has no orfs, yosys or verilator.
• buildifier is clean.

[gemini-code-assist]

Code Review

This pull request completely removes the bazel-orfs integration, including its associated tests (such as mock-array and gcd), configurations, and documentation, from the OpenROAD repository. It cleans up the MODULE.bazel and .bazelrc files by removing dependencies like rules_verilator, verilator, and orfs, while upgrading rules_bison and rules_flex to version 0.4 and adding rules_m4. The remaining documentation is updated to guide developers on how to run ORFS flows externally. The reviewer suggested clarifying the instructions for running the whittle.py script in docs/user/Bazel.md by specifying the correct working directory and paths.

[oharboe]

@maliberty @jhkim-pii Drinking my own cool-aid: drop heavy duty integration tests that nobody but the author will run in favour of extracted single concern working as intended tests.

[maliberty]

@gadfort FYI