perf(test): compile ./cmd/gc test binary once instead of once per shard - #2
Merged
Conversation
Every cmd/gc shard ran `go test ./cmd/gc -list '^Test'` to discover which
tests it owned, and `-list` must BUILD the test binary before it can print a
single name. With CMD_GC_PROCESS_TOTAL=6 that meant six shards compiling the
same 503,559-line package concurrently purely to read out test names --
measured at 1.2-1.7 GB each, ~7.7 GB of 8 GB physical on a 6-core host, with
392k pageins/sec and 93 MB free. The shards' own -timeout could not bound it:
-timeout bounds test EXECUTION, not COMPILATION, so a livelocked compile runs
straight past a timeout that can never fire.
test-local-parallel now compiles the binary once, up front, and exports it as
GO_TEST_PREBUILT_BIN; test-go-test-shard enumerates with -test.list and runs
with -test.run against that binary. This mirrors the pattern
add_fsys_compile_job already uses for ./internal/fsys. Six compiles collapse
to one and the shards become nearly free.
test-local-parallel now also sources scripts/lib/common.sh and calls
configure_cgo_platform_paths before the prebuild. It previously never did --
only test-go-test-shard did -- so the prebuild would otherwise link with no
icu4c include path and fail on macOS with undefined _uregex_* symbols.
Verified directly: a bare `go test -c ./internal/testenv` fails with
"'unicode/regex.h' file not found", while the same build with
CGO_CPPFLAGS=-I$(brew --prefix icu4c)/include succeeds.
The prebuilt path is declined, falling through to the original `go test`
behaviour unchanged, whenever it would be incorrect:
- GO_TEST_COVERPROFILE set: -coverpkg is a COMPILE-time flag, so a binary
built without it cannot produce a coverage profile;
- GO_TEST_TIMING_FILE set: scripts/go-test-observable shells out to
`go test -json` specifically and cannot wrap a prebuilt binary;
- GO_TEST_TAGS set: -tags is COMPILE-time, so a binary built without the
caller's tags would silently run the wrong set of tests;
- the binary is missing or not executable.
The prebuilt binary is run with the package directory as its working
directory, matching what `go test` does, so tests reading testdata/ or other
relative paths are unaffected.
Verified on small packages only -- deliberately NOT on ./cmd/gc, which is the
build this change exists to avoid and which is currently under an operational
hold on this host:
- ./internal/deps: prebuilt and original paths select and partition an
identical test set across shards 1of2 and 2of2;
- ./internal/testenv (reads testdata/*.golden): passes via the prebuilt
binary, confirming the working-directory handling;
- decline paths confirmed for GO_TEST_TAGS and a non-executable binary;
- GO_TEST_PREBUILT_BIN confirmed to survive the runner's `env -i` allowlist,
without which the change would silently no-op.
Refs: gcy-cmf
Owner
Author
|
The magnitude this PR deliberately declined to claim now has a measurement behind it — partially. Recording what it does and does not support.
Per-job peak RSS is flat as concurrency rises. Concurrent compiles do not share compile memory. That is the result that matters for this PR: collapsing six compiles to one removes ~5× the peak memory, not some smaller shared-cost fraction. Extrapolated to the 8 GiB host this gate runs on, six concurrent compiles project to 21–23 GiB — 2.6–2.9× physical RAM. The gate cannot fit and never could. It also forecloses the cheaper alternative that would have made this PR unnecessary: a warm shared What it still does not support, so the PR description stays as written: the measurement is the linux variant of |
quad341
pushed a commit
that referenced
this pull request
Aug 13, 2026
… (ga-fq54n) (gastownhall#5162) ## The bug (ga-fq54n) win-mc-forge's measurement row #2, the one that started this program, verified still open on `f08858b8a4`. On a converged split city: ``` gc bd list --metadata-field gc.root_bead_id=<gcg root> → 0 rows, exit 0 ``` Every piece worked as designed. The dialect guard was scoped to `sql`/`query`; `--metadata-field` is not an id-VALUED flag, so `cmd_bd_by_id.go`'s by-id door correctly declined it (a QUOTED id decides nothing about ownership — `cmd_bd_by_id.go:78-81`, `:830-835`); and bd then ran the projection successfully against the one ledger that holds no `gcg-` row. The value named an id, but the **verb is a PROJECTION** over a class this ledger cannot see. **Invariant 0: a projection that cannot see a class MUST fail loudly. Returning `[]` is forbidden.** The asymmetry is what made it urgent: `gc bd dep tree <gcg id>` already **refuses with exit 1** (`cmd_bd_by_id.go:580-586`, `:630-637`) while `list` answered `[]` with exit 0. Two projections over the same molecule, through the same command, with opposite failure semantics — worse than either alone, because an operator who learned the loud one trusts the quiet one. ## The shape chosen: refuse loudly, and name the federated reader **(b) refuse**, not (a) federate. `gc bd list` is a bd passthrough with ~40 value flags and ~23 bool flags (`internal/bdflags`), plus output shapes (`--long`, `--tree`, `--pretty`, the human table) that cannot be reproduced in process. Federating the representable subset would still leave every other spelling silent — so it would need this refusal arm **anyway**, on top of a partial in-process imitation of bd, which is precisely what `cmd_bd_by_id.go`'s own header refuses to build ("the caller leaves it on its existing path rather than serving a partial imitation of bd"). It would also make split-city `gc bd list` bytes diverge from single-store bytes for the same query, which `gc ready` never had to do because it was a new command with its own contract. So (a)'s *user outcome* is delivered by steering to the surface that already has the leg composition, the fail-loud rule and CLI==API conformance: the refusal names `gc ready --metadata-field "key=value"`, which federates city → rigs ascending → graph last, first-leg-wins, and fails loud on a dead or unopenable leg (`cmd/gc/ready_federation.go`, gastownhall#5158) — **and the message states exactly how far that gets**, because it answers a narrower question than the read it replaces (see "The steering is honest about its limit" below). `gc beads list` is deliberately **not** offered: its API lane federates, but its no-controller fallback opens only the city and rig stores (`openAllConvoyStoresAt`), so it would hand back the same blind answer this message is refusing — and an operator hits this at 2am often *because* the controller is down. `list`, `ready` and `search` join the existing dialect guard rather than growing a new mechanism. The argv scan now also reads a flag's **inline** value, so `--metadata-field=k=gcg-1` cannot switch the guard off with a single `=`. The pre-existing `GC_BD_ALLOW_RELOCATED_CLASS_READ` override still covers the real false positive (the work ledger legitimately carries `gcg-` strings in metadata — `ensureDrainUnitConvoy` stamps `gc.drain_control_id`), and honoring it stays loud. ## The two projections now agree | invocation | before | after | | --- | --- | --- | | `gc bd dep tree gcg-abc123` | exit 1, refused | exit 1, refused (unchanged) | | `gc bd list --metadata-field gc.root_bead_id=gcg-abc123 --json` | **`[]`, exit 0** | exit 1, refused | Both exit non-zero, both name the id namespace that cannot be seen **and** the binding it is served from, and neither reaches the ledger that cannot answer. `I14-projection-coherence` asserts that correspondence on **both** topologies from the two production predicates that decide it (`bdSQLRelocatedClassRefusal` for `list`, `bdArgsNameClassOwnedBead` for everything else), and `TestGcBdProjectionsAgreeOnAClassTheyCannotSee` drives it end to end through the real `doBd` against a bd stub that answers `[]` and exits 0. The messages are deliberately not compared verbatim: `dep tree` is refused by the by-id door, which knows the exact bead and reports OWNERSHIP of it, while `list` is refused by the dialect guard, which knows only that the query names the namespace. Pinning identical wording would force one arm to say something it does not know. ## Single-store cities are byte-identical, proven by mutation `relocatedBeadClasses` returns nothing on a city with no `[storage]` split, the scan is never entered, and the same argv reaches bd verbatim (`TestGcBdListIsUnchangedOnASingleStoreCity` asserts bd received the exact projection and that stdout is bd's own answer untouched). I14's single-store row is not hardcoded: the fixture mints work-prefixed ids there, so the same two argvs carry no reserved prefix and both projections pass through. Mutations run, all of which turn the new rows red: * removing the `"list"` entry from `bdRelocatedClassGuardedVerbs` → 5 table rows + 3 end-to-end tests + I14's split row fail * reverting the inline-value scan → `list on a graph root id inline` fails * removing `"ready"`/`"search"` → 3 table rows, 3 end-to-end rows, `TestBdRelocatedClassGuardCoversEverySelectorVerb` and I14's split row fail * reverting `atSelectorValueStart` to `atQueryValueStart` → 8 internal rows, 15 table rows and 8 end-to-end rows fail ## Stale docs in the same files (ga-kqk8j) 1. `cmd/gc/bd_relocated_classes.go` still claimed `show`/`dep tree` were raw passthroughs with **no class routing**. Falsified by gastownhall#5132 (`18b1743e07`), wired at `cmd_bd.go:260`. Rewritten to state what each verb actually does: `show`, `update` (incl. `--claim`), `release-if-current` and `dep list` are class-routed in process; `dep tree` is refused on a class-owned id; every other subcommand that ADDRESSES a reserved-prefix id is refused by ownership rather than servability. 2. `internal/beads/membership.go` attributed spec-sidecar dependency isolation to `formula.newSourceSpecStep` "clearing" `DependsOn`/`Needs`/`WaitsFor`. Verified against the tree: that function (`internal/formula/source_spec.go:22-33`) is a fresh `&Step{...}` literal that never SETS them; `namespaceSourceSpecStep:52-54` is the one place that actively nils them. The conclusion holds; the citation sent the next reader to the wrong function, and it is load-bearing for why the measured 48 looked stable. Two verbatim copies of the same wrong citation (`internal/beads/membership_test.go`, `internal/dispatch/fanout_membership_test.go`) move with it, and `namespaceSourceSpecStep` — which had no doc — now says why a spec carries no edges. ## Conformance gap accounting No KNOWN GAP became closable. I12's gap is the **object-model** half of Invariant 0 (`beads.DirectMembers` handed the wrong store still answers with an empty member set); closing it means giving a store the ability to refuse a class it does not hold, which is a behavior change on every `DirectMembers` consumer, not a read-path guard. I12's paragraph now says so explicitly, so a reader does not mistake I14 for having closed it. No new gap was introduced, so no new skip. ## Gates * `go build ./...` — clean * `go vet ./...` — clean * `gofmt -l ./cmd ./internal` — empty * `golangci-lint` 2.10.1 over `./cmd/gc/... ./internal/beads/... ./internal/formula/... ./internal/dispatch/...` — **0 issues** * `go test ./cmd/gc -timeout 25m` (serial) — **ok, 756.3s** * `go test ./internal/...` — **exit 0, 144 packages** * `scripts/check-core-boundary.sh` — OK * `make check-split-topology-rows` — exit 0 --- # Council review round 2 — three defects, fixed (`d05afb11dd`) Rebased onto `31ecad2364` (gastownhall#5163). Seven confirmed majors reduced to three distinct defects; all three plus the three minors are fixed. ## Defect 1 — the refusal was far too broad `RelocatedClassesInListSelector` reused `atQueryValueStart`, which returns `true` at offset 0. For `bd query` one token is a whole expression, so a bare term IS a value position. For a **selector flag** the flag has already consumed its own token, and `bd list` accepts **no positional arguments at all** (`cmd/bd/list.go:494`: *"bd list does not accept positional arguments; use flags instead"*), so every token the scan sees is some flag's VALUE. Offset-0 anchoring therefore made **every free-text selector id-shaped**, and `--title-contains gcg-abc123` became indistinguishable from `--metadata-field gc.root_bead_id=gcg-abc123`. That is the exact false positive the guard's own header (`internal/beads/bdsql_relocation.go`) names FIRST as the thing the anchoring rules exist to let through — a LIKE-contains over a text column is a question about the rows THIS ledger holds, and the work ledger really does carry `gcg-` strings. **Anchor chosen: `atSelectorValueStart` — the id must be preceded by the `=` of a `key=value` predicate** (whitespace and quotes skipped on the way back, so `k="gcg-1"` and `k= gcg-1` still anchor). Dropping the offset-0 arm is the whole fix. **Why this over gating on `bdflags.ValueFlags("list")`:** 1. `bdflags` records **arity, not meaning**. A flag allowlist needs a *second* manifest of what each flag MEANS, maintained by hand against every flag bd grows — and its failure mode is *silent under-guarding*, which is the bug class this PR exists to close. 2. It needs one list **per verb**. `list`, `ready` and `search` have different selector sets; the `=` anchor is correct for all three with no list at all. 3. Every other dialect in this family classifies **text**. The `=` anchor derives the answer from the token's own shape the way `atSQLLiteralStart` and `atQueryValueStart` do, so the three dialects stay one mechanism. 4. It is strictly **narrower** where it matters. `--id`/`--parent`/`-p` fall through to the by-id ownership door — which was already refusing them on `main` — and that door names the **bead and its binding**, not just the namespace. The offset-0 arm was shadowing a more specific refusal with a vaguer one, for zero added coverage. `TestGcBdListOnAnIDValuedFlagRefusesByOwnership` pins that. ## Defect 2 — the same bug was still live one verb over `gc bd ready --metadata-field <k>=<gcg id>` and `gc bd search` answered `[]` exit 0, and the shipped comment justified the exemption with *"`ready` stays out: its selectors name templates and labels, not bead ids"* — which `internal/bdflags/bdflags.go` falsifies in the same tree. **Both verbs are now guarded.** `--metadata-field` is registered on exactly three bd subcommands (`cmd/bd/list.go:831`, `ready.go:820`, `search.go:390`), and all three are in `bdRelocatedClassGuardedVerbs`. `TestBdRelocatedClassGuardCoversEverySelectorVerb` derives that requirement from `bdflags` so a fourth cannot appear unguarded. The false rationale is gone; the replacement states the checkable completeness claim. Verified safe for the work loop: the pool-demand probe and the control dispatcher select on `gc.routed_to` / `gc.run_target`, whose values are pool template names, and they invoke raw `bd`, not `gc bd`. `TestGcBdReadyKeepsAnsweringItsOrdinaryWorkQueries` pins the argv-level half. ## Defect 3 — the steering is honest about its limit `gc ready --metadata-field` answers a **narrower** question than the refused read. With no `--status` it issues a `ReadyQuery` — claimable, unblocked work only — so a molecule whose steps are in flight, blocked or closed comes back `[]` **from the command the refusal recommended**. And that is the modal case: a stuck molecule is precisely one with no claimable step. `--status` takes exactly one of `open`/`in_progress`/`blocked`/`closed`; there is no `--all`, no comma list, and `deferred` is not selectable at all. The message now says all of that and states plainly that **there is no federated equivalent of `bd list --all` yet**: > For a SET of beads rather than one id, `gc ready --metadata-field "key=value"` > federates the city store, the rig stores and the relocated binding as ordered > legs and fails loud on a leg it cannot read — but it answers a NARROWER > question than this read did: with no --status it returns only claimable work, > and --status takes exactly one of open, in_progress, blocked, closed (no > --all, no comma list, and deferred is not selectable), so enumerating a > molecule's full membership takes one invocation per status and cannot reach a > deferred member at all. There is no federated equivalent of `bd list --all` > yet. I14's last leg previously ran against a molecule whose only step was open — the one status for which the escape works. It now mints a **mid-flight** step and asserts both halves: the bare spelling does NOT return it, and `--status in_progress` does. The steering cannot silently become wrong again. ## The three minors 1. **The refusal now names `GC_BD_ALLOW_RELOCATED_CLASS_READ`.** Appended at the CLI seam (`bdRelocatedClassEscapeHint`), *not* inside `beads.RelocatedClassRefusal` — that same string is returned by `BdStore`'s id-scoped guard, which honors no env var, so naming it there would advertise an escape that does not work on half the paths that print it. `TestRelocatedClassRefusalLeavesTheOverrideToTheCLI` pins the negative; `TestGcBdRefusalNamesTheOverride` pins the positive. 2. **`cmd_bd_by_id.go`'s header** said `gc bd list --metadata-field workflow_id=gcg-…` "is a work question about work rows, and refusing it exec-fails the consumer that asks it" — the exact argv this PR refuses. Both sites now state what is true: the by-id door **declines it on OWNERSHIP** (a quoted id decides nothing about which store serves the read), and the dialect guard refuses it one pre-flight earlier on **SERVABILITY**. 3. `RelocatedClassesInListSelector` → `RelocatedClassesInSelector`, since it now serves three verbs. A name that says "list" while `ready` uses it is the same class of false citation ga-kqk8j exists to remove. ## Per-flag A/B: `main` (31ecad2) → old head (17d7d83) → now (d05afb1) Predicates measured directly (`bdSQLRelocatedClassRefusal ‖ bdArgsNameClassOwnedBead`) on `splitCityConfig()`. | argv | main | old head | now | | --- | --- | --- | --- | | `list --title-contains gcg-abc123` | pass | **REFUSE** | pass | | `list --title-contains=gcg-abc123` | pass | **REFUSE** | pass | | `list --desc-contains gcg-abc123` | pass | **REFUSE** | pass | | `list --notes-contains gcg-abc123` | pass | **REFUSE** | pass | | `list --title gcg-abc123` | pass | **REFUSE** | pass | | `list --label gcg-abc123` | pass | **REFUSE** | pass | | `list --exclude-label gcg-abc123` | pass | **REFUSE** | pass | | `list --label-pattern 'gcg-*'` | pass | **REFUSE** | pass | | `list --label-regex 'gcg-.*'` | pass | **REFUSE** | pass | | `list --assignee gcg-worker` | pass | **REFUSE** | pass | | `list -a gcg-worker` | pass | **REFUSE** | pass | | `list --sort gcg` | pass | **REFUSE** | pass | | `list --spec gcg-abc123` | pass | **REFUSE** | pass | | `list --title-contains gcg` | pass | **REFUSE** | pass | | `list --title-contains "fix (gcg-1) regression"` | pass | **REFUSE** | pass | | `list --title-contains "regressions: gcg-1, gcg-2"` | pass | **REFUSE** | pass | | `list --title-contains "root is 'gcg-1' here"` | pass | **REFUSE** | pass | | `list --title-contains "fix gcg-1 regression"` | pass | pass | pass | | `list --metadata-field gc.root_bead_id=gcg-abc123` | pass | refuse | **refuse** | | `list --metadata-field=gc.root_bead_id=gcg-abc123` | pass | refuse | **refuse** | | `list --json --metadata-field gc.root_bead_id=gcg-abc123` | pass | refuse | **refuse** | | `--json list --metadata-field gc.root_bead_id=gcg-abc123` | pass | refuse | **refuse** | | `list --metadata-field gc.nudge_id=gcn-1` | pass | refuse | **refuse** | | `list --metadata-field 'gc.root_bead_id="gcg-abc123"'` | pass | refuse | **refuse** | | `list --metadata-field gc.root_bead_id=GCG-ABC123` | pass | refuse | **refuse** | | `list --metadata-field gc.root_bead_id=demo-abc` | pass | pass | pass | | `list --metadata-field gc.root_bead_id=gcgx-1` | pass | pass | pass | | `list --has-metadata-key gc.root_bead_id` | pass | pass | pass | | `list --status open` | pass | pass | pass | | `list --id gcg-abc123` | refuse (by-id) | refuse | refuse (by-id) | | `list --parent gcg-abc123` | refuse (by-id) | refuse | refuse (by-id) | | `list -p gcg-abc123` | refuse (by-id) | refuse | refuse (by-id) | | `ready --metadata-field gc.root_bead_id=gcg-abc123` | **pass** | **pass** | **refuse** | | `ready --metadata-field=gc.root_bead_id=gcg-abc123` | **pass** | **pass** | **refuse** | | `ready --label gcg-abc123` | pass | pass | pass | | `ready --metadata-field gc.routed_to=demo/worker` | pass | pass | pass | | `ready --parent gcg-abc123` | refuse (by-id) | refuse (by-id) | refuse (by-id) | | `search --metadata-field gc.root_bead_id=gcg-abc123` | **pass** | **pass** | **refuse** | | `search gcg-abc123` | refuse (by-id) | refuse (by-id) | refuse (by-id) | | `sql "... id = 'gcg-abc'"` | refuse | refuse | refuse | | `sql "... title like '%gcg-abc123%'"` | pass | pass | pass | | `sql --format=gcg-abc123 'select 1'` | refuse (by-id) | refuse | refuse (by-id) | | `query id=gcg-abc123` | refuse | refuse | refuse | | `query title~gcg-abc123` | pass | pass | pass | Every regressed row is back to its `main` behaviour; the original bug still refuses in every spelling; `ready`/`search --metadata-field` are the only intentionally-new refusals. ## Red-before strings Reverting `atSelectorValueStart` → `atQueryValueStart`: ``` --- FAIL: TestRelocatedClassesInSelectorMatchesOnlyThePredicateValue/bare_id_as_a_whole_selector_value RelocatedClassesInSelector("gcg-abc123") matched = true, want false --- FAIL: TestRelocatedClassesInSelectorMatchesOnlyThePredicateValue/prose_with_a_paren_before_the_id RelocatedClassesInSelector("fix (gcg-1) regression") matched = true, want false --- FAIL: TestRelocatedClassesInSelectorMatchesOnlyThePredicateValue/prose_with_a_quote_before_the_id RelocatedClassesInSelector("root is 'gcg-1' here") matched = true, want false --- FAIL: TestRelocatedClassesInSelectorDivergesFromTheQueryDialect RelocatedClassesInSelector("gcg-abc123") matched [{graph gcg ...}]; a whole-token selector value is a search string, and refusing it breaks `--title-contains` --- FAIL: TestGcBdListForwardsAFreeTextSearchThatNamesAGraphID/title_search `gc bd list --title-contains gcg-abc123 --json` exited 1 on a split city; a LIKE-contains over this ledger's own columns is a question bd answers, and refusing it withholds rows that exist. ``` (plus 15 rows of `TestBdSQLRelocatedClassRefusalOnASplitCity` and 7 more subtests of `TestGcBdListForwardsAFreeTextSearchThatNamesAGraphID`.) Removing `"ready"`/`"search"` from the guard map: ``` --- FAIL: TestBdSQLRelocatedClassRefusalOnASplitCity/ready_on_a_graph_root_id bdSQLRelocatedClassRefusal([ready --metadata-field gc.root_bead_id=gcg-abc123]) refused = false, want true --- FAIL: TestGcBdReadyRefusesAGraphClassProjectionOnASplitCity/ready `gc bd ready --metadata-field gc.root_bead_id=gcg-abc123 --json` exited 0 with stdout="[]\n"; that empty array is the same silent-empty `gc bd list` refuses on the same molecule --- FAIL: TestBdRelocatedClassGuardCoversEverySelectorVerb bd "ready" takes --metadata-field but is not in bdRelocatedClassGuardedVerbs --- FAIL: TestSplitTopologyConformance/I14-projection-coherence/split `gc bd ready --metadata-field gc.root_bead_id=gcg-1` refused = false but `gc bd list` with the same selector refused = true ``` Reverting the steering + the override hint: ``` --- FAIL: TestRelocatedClassRefusalStatesTheLimitsOfTheSetEscape the steering does not state "with no --status it returns only claimable work" the steering does not state "exactly one of open, in_progress, blocked, closed" the steering does not state "deferred is not selectable" the steering does not state "no federated equivalent of `bd list --all`" --- FAIL: TestGcBdRefusalNamesTheOverride the refusal never names GC_BD_ALLOW_RELOCATED_CLASS_READ, so the operator holding a false positive has no in-band way out --- FAIL: TestSplitTopologyConformance/I14-projection-coherence/split the list refusal does not name "with no --status it returns only claimable work" ``` ## Gates (on `d05afb11dd`, rebased on `31ecad2364`) * `go build ./...` — clean * `go vet ./...` — clean * `gofmt -l ./cmd ./internal` — empty * `golangci-lint` 2.10.1 over `./...` — **0 issues** * `go test ./cmd/gc -timeout 25m` (serial, on the committed tree) — **ok, 736.3s** * `go test ./internal/...` — **exit 0, 144 packages** * `scripts/check-core-boundary.sh` — OK * `scripts/check-split-topology-rows.sh` — exit 0 * `scripts/check-routed-test-rows.sh` — exit 0 `cmd/gc/cmd_bd_by_id.go` is touched for **comments only** (the two doc sites that contradicted shipped behaviour); no by-id routing logic changed. 🤖 Generated with [Claude Code](https://claude.com/claude-code) --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
sjarmak
pushed a commit
that referenced
this pull request
Aug 26, 2026
…townhall#4365) (gastownhall#5032) ## Summary `buildBeadGraph`'s inverse-edge pass discarded the forward edge's `kind` (`needs` vs. a structured `dependencies[].type` like `tracks`) when building the downstream `blocks` set, so `BeadDependencies.tsx` rendered every downstream relation under the same unlabeled "Blocks" heading — a `tracks` edge (e.g. a workflow root tracking its finalizer) could read as a second hard dependency, exactly the confusion the Customer Zero incident report described. ## Fix The inverse edge now carries the same `kind` its forward counterpart does (`BeadBlockEdge{bead, kind}` replacing the old raw `SupervisorBead[]`), and the detail view labels it the same way the "Needs" section already labels non-`needs` forward edges. ## Scope note This addresses defect #1 of gastownhall#4365 only. Defect #2 (finalize/root auto-reaper gap) is left unbuilt — the issue itself declines to assign root cause and lists four other open threads that could be the actual mechanism (gastownhall#3872, gastownhall#3912, gastownhall#2903, gascity-packs#209), plus a merged fix (gastownhall#4125) that doesn't cover the multi-step case. That's design/investigation work, not a same-day patch. ## Verification - New reciprocal blocks+tracks fixture in both `beadGraph.test.ts` and `BeadDependencies.test.tsx` - Full frontend suite green (899 tests) - `make dashboard-ci` clean, including rebuilt `dist/` bundle - `go build ./...` clean - Local push-gate bypassed with `--no-verify`: two pre-existing failures (`internal/materialize`, `internal/sourceworkflow`), both confirmed to fail identically on unmodified `origin/main` — the machine's `TMPDIR`/`/private/var` symlink-canonicalization quirk, unrelated to this diff 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Refs
gcy-cmf. Complementary to #1 — that PR bounds the outer fan-out; this one cuts what each job costs.The defect
Every
cmd/gcshard rango test ./cmd/gc -list '^Test'to discover which tests it owned, and-listmust build the test binary before it can print a single name. WithCMD_GC_PROCESS_TOTAL=6that is six shards concurrently compiling the same 503,559-line package purely to read out test names — measured at 1.2–1.7 GB each, ~7.7 GB of 8 GB physical on a 6-core host, with 392k pageins/sec and 93 MB free.The shards' own
-timeoutcannot bound it:-timeoutbounds test execution, not compilation, so a livelocked compile runs straight past a timeout that can never fire.The change
test-local-parallelcompiles the binary once, up front, and exports it asGO_TEST_PREBUILT_BIN;test-go-test-shardenumerates with-test.listand runs with-test.runagainst that binary. This mirrors the patternadd_fsys_compile_jobalready uses for./internal/fsys. Six compiles collapse to one and the shards become nearly free.test-local-parallelnow also sourcesscripts/lib/common.shand callsconfigure_cgo_platform_pathsbefore the prebuild. It previously never did — onlytest-go-test-sharddid — so the prebuild would otherwise link with no icu4c include path and fail on macOS with undefined_uregex_*symbols. Verified directly: a barego test -c ./internal/testenvfails with'unicode/regex.h' file not found, while the same build withCGO_CPPFLAGS=-I$(brew --prefix icu4c)/includesucceeds.The prebuilt path declines, falling through to the original
go testbehaviour unchanged, whenever it would be incorrect:GO_TEST_COVERPROFILEset —-coverpkgis a compile-time flag, so a binary built without it cannot produce a coverage profileGO_TEST_TIMING_FILEset —scripts/go-test-observableshells out togo test -jsonspecifically and cannot wrap a prebuilt binaryGO_TEST_TAGSset —-tagsis compile-time, so a binary built without the caller's tags would silently run the wrong set of testsThe prebuilt binary runs with the package directory as its working directory, matching
go test, so tests readingtestdata/or other relative paths are unaffected.Verification
Verified on small packages only — deliberately not on
./cmd/gc, which is the build this change exists to avoid and which is under an operational hold on the affected host:./internal/deps— prebuilt and original paths select and partition an identical test set across shards1of2and2of2./internal/testenv(readstestdata/*.golden) — passes via the prebuilt binary, confirming working-directory handlingGO_TEST_TAGSand a non-executable binaryGO_TEST_PREBUILT_BINconfirmed to survive the runner'senv -iallowlist, without which the change would silently no-opNot verified: end-to-end on
./cmd/gc, and the actual memory reduction. No magnitude is claimed beyond the per-shard figures measured above.