test/goroot: cover compile-only, errorcheck, and remaining runnable Go 1.26 cases

[cpunion]

Status

Complete Go 1.26 GOROOT coverage is implemented on the current-main candidate and validated with Go 1.26.5 on both target platforms.

• Integration base: xgo-dev/main at 112181e06
• Current-main coverage branch: codex/goroot-go126-coverage-main at da9f1e054
• Draft PR: test(goroot): enable complete Go 1.26 coverage #2065
• Preserved merged-PR comparison: codex/goroot-go126-coverage at d40dddb66 (closed PR test(goroot): enable complete Go 1.26 coverage #2064)
• Toolchain: Go 1.26.5
• Target platforms: darwin/arm64, linux/amd64
• Runner mode: coverage
• Unexpected unclassified failures: 0
• Go 1.26 host skips on both target platforms: 0

The runner discovers every active compile, errorcheck, errorcheckandrundir, run, runoutput, rundir, runindir, buildrun, and buildrundir case in the configured GOROOT test directories. Compile-only and diagnostic tests execute rather than being inferred or omitted.

Current-main Go 1.26.5 inventory and results

pass and fail describe the observed case result. A diagnostic test passes only when LLGo's diagnostics satisfy its // ERROR expectations. Xfails and failed flakes remain executed failures in this inventory.

The table starts from the complete current-main sweep and applies the individually verified former-skip delta: six cases moved from skip to pass, while the executed Darwin chanlinear.go result moved from skip to failed flake. It is an exact inventory/result reconciliation, not a second full sweep.

Directive	Total	Pass	Fail	Skip
buildrun	1	1	0	0
buildrundir	2	0	2	0
compile	510	499	11	0
errorcheck	650	472	178	0
errorcheckandrundir	5	1	4	0
run	1,030	982	48	0
rundir	116	110	6	0
runindir	9	6	3	0
runoutput	22	20	2	0
Total	2,345	2,091	254	0

Overall classification:

• 2,087 other cases: passed
• 246 xfails: failed as expected
• 7 flakes: 4 passed and 3 failed
• 5 resource-guard stops: bounded hard failures from the complete current-main sweep
• 0 host skips for Go 1.26 on either target platform
• 0 unexpected unclassified failures
• 2,345 / 2,345 represented and executed (100%) across the complete sweep plus targeted former-skip verification

Failure conservation is 246 xfails + 3 failed flakes + 5 resource stops = 254 observed failures. Xfails remain part of coverage: each case is compiled, run, or diagnostic-matched before classification.

Final current-main flake outcomes on darwin/arm64:

• Passed: rotate0.go, fixedbugs/issue79186.go, fixedbugs/issue43942.go, fixedbugs/issue48898.go
• Failed: goprint.go, stack.go, chanlinear.go

Current main vs merged-PR comparison

The merged-PR comparison is preserved as data; it has not been rebased over the current-main candidate. Delta is merged PRs minus current main.

Directive	Total	Merged pass	Main pass	Delta pass	Merged fail	Main fail	Delta fail	Merged skip	Main skip
buildrun	1	1	1	0	0	0	0	0	0
buildrundir	2	0	0	0	2	2	0	0	0
compile	510	499	499	0	11	11	0	0	0
errorcheck	650	470	472	-2	180	178	+2	0	0
errorcheckandrundir	5	1	1	0	4	4	0	0	0
run	1,030	984	982	+2	39	48	-9	7	0
rundir	116	112	110	+2	4	6	-2	0	0
runindir	9	8	6	+2	1	3	-2	0	0
runoutput	22	21	20	+1	1	2	-1	0	0
Total	2,345	2,096	2,091	+5	242	254	-12	7	0

The two extra merged-branch errorcheck failures are fixedbugs/bug299.go and fixedbugs/issue5089.go, both introduced by PR #2031. Detailed attribution and the complete failure-reason taxonomy are in this issue update.

Former host-skip validation

Commit d6cfed406 removes all Go 1.26 host skips while preserving the Go 1.24/1.25 protections. Unversioned Darwin rules were narrowed to Go 1.24/1.25 instead of being deleted globally.

Case	darwin/arm64	linux/amd64
ddd.go	pass	pass
gcgort.go	pass	pass
fixedbugs/issue25897b.go	pass	pass
fixedbugs/issue27695.go	pass	pass
fixedbugs/issue39541.go	pass	pass
chan/goroutines.go	pass	pass
chanlinear.go	executed; known timing flake	pass

Darwin local validation selected all 7 cases and reported 7 passing subtests with no skips. Six passed cleanly; chanlinear.go reproduced the LLGo chanSelect linearity timing failure and is now an executed flake instead of a host skip. Peak process-group RSS was about 921 MiB under a 15 GiB hard limit.

Linux was validated on a native GitHub ubuntu-latest amd64 runner with Go 1.26.5. All 7 cases passed cleanly with no skips; peak RSS was about 982 MiB under a 4 GiB hard limit: Actions job. The optional hosted macOS duplicate was canceled after waiting 25 minutes for a runner; local darwin/arm64 validation is the authoritative Darwin result.

Go 1.26.5 delta

Go 1.26.0 contained 2,332 matching cases. Go 1.26.5 contains 2,345, adding 13:

Directive	Go 1.26.0	Go 1.26.5	Delta
compile	504	510	+6
run	1,023	1,030	+7
Other directives	805	805	0
Total	2,332	2,345	+13

All 13 Go 1.26.5 additions passed LLGo validation.

Failure classification detail

All 254 current-main failures have a technical classification:

• buildrundir: 2 Go assembler asmhdr/symabis and return-jump ABI gaps
• compile: 10 unsupported gc/backend flags and 1 RSS resource stop
• errorcheck: 177 diagnostic xfails and 1 RSS resource stop
• errorcheckandrundir: 4 diagnostic xfails; the mandatory build/run stage is still executed
• rundir / runindir: 9 expected directory-case failures
• run: 44 xfails, 3 failed flakes, and 1 resource stop
• runoutput: 2 resource stops

The detailed taxonomy separates gc-only compiler analysis/backend/runtime diagnostics from LLGo frontend wording/recovery/position differences and groups runnable failures by GC/liveness, panic/recover, stack metadata, compile/link, interface/reflection, and scheduler behavior. See the linked taxonomy comment above for counts and case boundaries.

Implemented coverage

• Added coverage discovery mode with recursive nested-case discovery.
• Added -directives, -list-cases, and -list-case-paths for reproducible inventory and sharding.
• Added llgo tool compile with Go compiler-compatible flag names and value forms used by cmd/internal/testdir, including -c, -C, -e, -N, -l, -lang, -complete, -d, -p, -importcfg, -D, -I, and -o.
• Propagated primary-package language versions and per-file versions through package loading.
• Added compile-only handling for declarations without bodies.
• Implemented compile and errorcheck execution with upstream-style compiler options and // ERROR matching.
• Implemented errorcheckandrundir as diagnostic checking plus a mandatory build/run stage.
• Implemented companion-directory staging for rundir, runindir, and buildrundir, including _test.go, relative imports, assembly placeholders, and nested module layouts.
• Added recursive nested run cases and official run working-directory/environment behavior.
• Preserved non-zero process exits and timeout/resource errors through xfail classification.

Resource controls

The complete current-main sweep used serial execution with GOMAXPROCS=2, go test -p=1, and a 1,536 MiB process-group RSS hard limit. It intentionally recorded five bounded resource stops rather than allowing xfail entries to hide them.

The former-skip validation used:

• Local Darwin: GOMEMLIMIT=12GiB, 15 GiB RSS hard limit, 10 ms RSS polling, minimum 40% free memory, and minimum 1,536 MiB free swap.
• Native Linux CI: GOMEMLIMIT=3GiB, 4 GiB RSS hard limit, 25 ms RSS polling, and GOMAXPROCS=2.

The runner:

• checks physical-memory and swap thresholds before and during each command;
• samples total RSS for the complete Unix process group;
• kills the process group when the hard RSS limit is exceeded;
• reports high-water RSS warnings;
• treats resource-guard stops as hard failures that xfail entries cannot hide;
• defaults the helper script to GOMAXPROCS=2 and go test -p=1.

These guards prevent recurrence of the unbounded LLGo coverage build that previously reached approximately 32.5 GiB RSS and exhausted swap.

Reproduction

Inventory only:

GOMAXPROCS=2 go test -p=1 ./test/goroot -run '^TestGoRootRunCases$' -count=1 -args \
  -goroot "$(go env GOROOT)" \
  -directive-mode coverage \
  -list-cases

Resource-constrained execution:

GOMAXPROCS=2 GOMEMLIMIT=12GiB go test -p=1 ./test/goroot -run '^TestGoRootRunCases$' -count=1 -timeout=180m -args \
  -goroot "$(go env GOROOT)" \
  -directive-mode coverage \
  -max-rss-mib 15360 \
  -rss-warn-mib 512 \
  -min-memory-free-percent 40 \
  -min-swap-free-mib 1536 \
  -rss-poll 10ms

Acceptance criteria

• All 510 active Go 1.26.5 compile cases are discovered and classified.
• All 650 errorcheck cases execute diagnostic matching and are classified.
• All 5 errorcheckandrundir cases execute diagnostic and run stages.
• All 1,185 runnable cases are discovered and classified.
• All 2,345 active cases are represented in the inventory and result totals.
• Every Go 1.26 case has been executed; no target-platform host skip remains.
• Former host skips are validated on darwin/arm64 and linux/amd64 under bounded resources.
• Compile-only success does not require a linked executable.
• Resource use is bounded independently of xfail handling.

Pull request structure and CI status

The prerequisite implementation now uses two coherent review units:

• Draft cl: handle SSA field values in unsafe.Offsetof #2068: independent unsafe.Offsetof SSA-field fix with a real test/go regression (3 files, 67 additions).
• Draft cmd: add GOROOT compile-test adapter #2066: the single Go compiler parameter compatibility PR, including loader/build propagation, llgo tool compile, internal tests, and native go tool compile vs LLGo test/go coverage (13 files, 1,080 additions, 14 deletions).

The earlier fine-grained PRs #2069, #2070, and #2071 are closed as superseded by #2066. Their branches remain preserved. Merge #2068 and #2066 independently, then rebase draft #2065 onto main.

The previous CLI-only #2066 head is preserved as codex/go-tool-compile-cli-only; the consolidated implementation is mirrored at codex/go-tool-compile-action-cli-full.

Bounded local validation passes with Go 1.24.11 and Go 1.26.5 for both ordinary Go tests and real llgo test ./test/go runs. The #2068 regression fails on unmodified main with the expected unsafe.Offsetof panic. The #2066 compatibility test found and now covers multiline -C driver diagnostics. #2066 and #2068 each completed with 47 passing checks, 1 expected release skip, and 100.00% Codecov patch coverage.

Independent #2072 remains the select-output CI stability fix; its prior complete run had 47 passing checks, 1 expected release skip, and no product coverage change.

[cpunion]

Draft implementation: #2064. The PR is based on the merged stage5 integration branch, updated to the latest main, and includes the complete Go 1.26.5 statistics and resource-guard validation documented above.

[cpunion]

Go 1.26.5 full coverage comparison: merged PRs vs current main

Two independent Go 1.26.5 sweeps are now available on darwin/arm64:

• Merged PRs: codex/goroot-go126-coverage at d40dddb66, preserved in closed comparison PR test(goroot): enable complete Go 1.26 coverage #2064.
• Current main only: codex/goroot-go126-coverage-main at 8b4b9248e, based on xgo-dev/llgo main at 112181e06 plus only the coverage implementation.

Delta below is merged PRs minus current main.

Directive	Total	Merged pass	Main pass	Delta pass	Merged fail	Main fail	Delta fail	Merged skip	Main skip
buildrun	1	1	1	0	0	0	0	0	0
buildrundir	2	0	0	0	2	2	0	0	0
compile	510	499	499	0	11	11	0	0	0
errorcheck	650	470	472	-2	180	178	+2	0	0
errorcheckandrundir	5	1	1	0	4	4	0	0	0
run	1,030	984	976	+8	39	47	-8	7	7
rundir	116	112	110	+2	4	6	-2	0	0
runindir	9	8	6	+2	1	3	-2	0	0
runoutput	22	21	20	+1	1	2	-1	0	0
Total	2,345	2,096	2,085	+11	242	253	-11	7	7

Failure classification comparison

Classification	Merged PRs	Current main
Expected failures / xfails	238	246
Flaky cases that failed	4	2
Resource-guard stops	0	5
Unexpected unclassified failures	0	0
Observed failures	242	253

Both inventories were complete: 2,345 mapped / 2,345 observed, with 2,338 executed cases and the same 7 host-safety skips.

The merged result had 2 flaky passes and 4 flaky failures. The main result had 4 flaky passes and 2 flaky failures (goprint.go and stack.go).

The main-only sweep had five reproducible resource stops:

• compile/fixedbugs/issue15141.go
• errorcheck/fixedbugs/issue20780.go
• runoutput/fixedbugs/bug449.go
• runoutput/rangegen.go (the Go baseline generated build exceeded the cap)
• run/recover.go

Both runs used serial execution, GOMAXPROCS=2, go test -p=1, and a 1536 MiB process-group RSS hard limit. The main-only rerun tightened RSS polling from 50 ms to 20 ms, so the resource-stop difference should not be treated as a pure code delta.

The main-only branch additionally removed two stale Go 1.26 errorcheck xfails, narrowed 24 stale darwin xfails to Go 1.24/1.25 after individual and combined Go 1.26.5 pass verification, and added five stable Go 1.26 xfails found by expanded coverage.

Both remote branches remain preserved; neither was overwritten.

Attribution of the two errorcheck regressions

The merged branch has two additional errorcheck failures compared with current main:

• fixedbugs/bug299.go
• fixedbugs/issue5089.go

Both are introduced by PR #2031, commit 2173cea62 (ssa,runtime: gc-conformant type-assertion panics and go:nointerface support). The PR added FuncDecl processing in cl.ParsePkgSyntax, which calls astFuncName; these negative tests contain receiver ASTs that reach recvTypeName / astFuncName and panic with unreachable before the expected diagnostics are emitted.

This was verified on the final merged tree: removing only the new ParsePkgSyntax -> astFuncName path and rebuilding makes both cases pass with xfail handling disabled. No other stage5 PR needs to be removed for the two diagnostics to recover.

Failure-reason taxonomy for the #2065 current-main baseline

All 253 observed failures are assigned to a technical cause below. Status conservation remains 246 expected failures + 2 failed flakes + 5 resource stops = 253.

The 33 run xfails that previously had generic reasons such as "current/latest main failure" were rerun serially with only xfail classification disabled. Timeout overrides, flake handling, host skips, and all memory/RSS guards remained enabled, so the categories below use their actual failure output rather than file-name inference.

Failure status by directive

Directive	Fail	Xfail	Flaky failure	Resource stop
buildrundir	2	2	0	0
compile	11	10	0	1
errorcheck	178	177	0	1
errorcheckandrundir	4	4	0	0
run	47	44	2	1
rundir	6	6	0	0
runindir	3	3	0	0
runoutput	2	0	0	2
Total	253	246	2	5

buildrundir (2)

Cause	Count
Go assembler asmhdr/symabis pipeline and return-jump ABI behavior not implemented	2

compile (11)

Cause	Count
Unsupported gc-only backend/debug flags	10
RSS resource guard	1

The 10 flag cases are: -d=libfuzzer (3), -d=checkptr (2), -d=ssa/check/seed (2), -B bounds checking (1), -d=softfloat (1), and -dynlink (1). The resource stop is fixedbugs/issue15141.go.

errorcheck (178)

Cause	Count
gc escape/inlining/compiler-analysis diagnostics	64
Parser/recovery diagnostic differences	45
gc runtime pragmas, liveness, and write-barrier diagnostics	20
gc SSA/backend optimization diagnostics	18
Type-checker semantic wording/source-position differences	13
Scanner/lexical diagnostic differences	9
Loader/import/module diagnostic differences	8
RSS resource guard	1
Total	178

Thus 102/178 are gc-specific compiler-analysis/backend/runtime diagnostic expectations, 75/178 are LLGo-vs-Go frontend diagnostic wording/recovery/position differences, and 1/178 is the bounded resource stop fixedbugs/issue20780.go.

errorcheckandrundir (4)

Cause	Count
gc escape/inlining diagnostics	3
gc SSA intrinsic diagnostics	1

These four fail only the compiler-diagnostic expectation; the runner still executes their mandatory build/run stage.

Combined run* family (58)

This combines run (47), rundir (6), runindir (3), and runoutput (2).

Cause	Count
GC/finalizer/liveness/heap profiling	12
Panic/recover/defer/nil semantics	12
Stack/caller/traceback and line metadata	12
Compile/link/init/fieldtrack/codegen	9
Interface/type identity/reflection	7
Scheduler/goroutine/stack execution	3
Resource guard	3
Total	58

Notable boundaries:

• The stack/metadata group includes incorrect panic lines, missing callers/traceback markers, defer caller lines, CallersFrames metadata, and bad-PC traceback lookup.
• The GC group includes heap-sampling attribution, finalizer timing, unsafe-uintptr liveness, stack-object liveness, and GC progress.
• The interface group includes cross-package type identity/itab behavior, go:nointerface, reflection field values, and interface-conversion diagnostic identity.
• The compile/link group includes missing inittask, range/generic builds that produce no binary, missing maymorestack instrumentation, bodyless-symbol reachability, package init order, and fieldtrack metadata.
• The two failed flakes are goprint.go and stack.go in the scheduler/stack group.
• The three run* resource stops are recover.go, fixedbugs/bug449.go, and rangegen.go; rangegen.go exceeded the cap in the Go baseline generated build.

[cpunion]

Go 1.26.5 host-skip closure on current-main coverage

Draft PR #2065 now includes commit d6cfed406, which enables the previously skipped Go 1.26 run cases while preserving the Go 1.24/1.25 host-safety rules.

Cross-platform targeted validation

Case	darwin/arm64	linux/amd64
ddd.go	pass	pass
gcgort.go	pass	pass
fixedbugs/issue25897b.go	pass	pass
fixedbugs/issue27695.go	pass	pass
fixedbugs/issue39541.go	pass	pass
chan/goroutines.go	pass	pass
chanlinear.go	executed; known timing flake	pass

Darwin was run locally on Go 1.26.5 with serial execution, GOMAXPROCS=2, go test -p=1, GOMEMLIMIT=12GiB, a 15 GiB process-group RSS hard limit, 10 ms RSS polling, and system-memory/swap guards. The configured suite selected all 7 cases, reported 7 passing subtests with no skips, and observed a maximum process-group RSS of about 921 MiB. Six cases passed cleanly; chanlinear.go reproduced the LLGo chanSelect linearity timing failure and was handled as an executed flake rather than a host skip.

Linux was verified on the native GitHub ubuntu-latest amd64 runner with Go 1.26.5, GOMAXPROCS=2, GOMEMLIMIT=3GiB, and a stricter 4 GiB RSS hard limit. All 7 cases passed cleanly, no host skip was applied, and peak RSS was about 982 MiB: Actions job. The optional hosted macOS duplicate was canceled after waiting 25 minutes for a runner; the same Darwin configuration passed locally under the 15 GiB guard.

Expectation changes

• Go 1.26 matching host skips: darwin/arm64 7 -> 0, linux/amd64 2 -> 0.
• The unversioned Darwin rules for ddd.go, gcgort.go, and the three fixedbugs cases were narrowed to Go 1.24/1.25 instead of being deleted globally.
• Go 1.24/1.25 platform-specific skips remain unchanged in effect.
• Go 1.26 chanlinear.go on Darwin is now an executed timing flake, because it is resource-safe but can fail the official 3x linearity threshold. Linux remains an ordinary passing case.

Revised current-main statistics

These totals are the previous complete current-main inventory plus the individually verified 7-case delta; this is not a new full sweep. The six clean passes move from skip to pass, while the executed Darwin chanlinear.go result moves from skip to flaky failure.

Directive	Total	Pass	Fail	Skip
run	1,030	982	48	0
all other directives	1,315	1,109	206	0
Total	2,345	2,091	254	0

Failure conservation for the revised current-main result is 246 xfails + 3 failed flakes + 5 resource stops = 254 observed failures. Coverage is now 2,345 mapped / 2,345 observed / 2,345 executed, with no Go 1.26 host skip on either target platform.

The preserved merged-PR comparison remains unchanged at 2,096 pass / 242 fail / 7 skip. Relative to revised current main, its total delta is now +5 pass / -12 fail / +7 skip; for the run directive alone it is +2 pass / -9 fail / +7 skip.

[cpunion]

Go 1.26.5 full baseline on the latest main (Darwin + Linux)

A fresh two-platform sweep has completed against the current upstream main.

• Upstream main: 41a336e (includes merged build: schedule package initialization in Go order #2254)
• Coverage-run head: 63db0fd (the upstream commit plus the test-only workflow)
• Actions run: 31295941212
• Matrix: Go 1.26.5, darwin/arm64 + linux/amd64, directive-mode coverage, 4 shards per platform
• Runner: GOMAXPROCS=2, go test -p=1, 180-minute test timeout; 512 MiB minimum free swap on Linux and no free-swap threshold on macOS

The run concluded as failure because coverage mode deliberately fails stale expectations and unexpected failures. All eight jobs nevertheless completed their full shards; there was no canceled or truncated shard.

Complete observed-status inventory

Linux has 21 more platform-selected observations than Darwin. The union contains 2,368 distinct case paths.

Observed status	Darwin occurrences	Linux occurrences	Both platforms	Distinct case paths
Ordinary pass	2,145	2,125	4,270	2,157
Actual xfail	42	41	83	46
Not applicable	116	112	228	116
Stale xfail: case actually succeeded	22	43	65	44
Stale not-applicable: case actually succeeded	1	1	2	1
Known flake passed	5	7	12	10
Known flake failed	3	1	4	4
Resource guard stopped case	0	2	2	2
Unexpected, unclassified failure	11	34	45	37
Total	2,345	2,366	4,711	2,368

The distinct-path column is status-specific and is therefore not additive: a platform-specific xfail can be an ordinary or unexpected result on the other platform. The occurrence columns are mutually exclusive and sum exactly to 4,711.

The two resource stops were Linux-only:

• run/fixedbugs/issue20780b.go
• runoutput/fixedbugs/bug449.go

macOS completed all 2,345 observations without a resource stop, confirming that disabling the inappropriate free-swap threshold avoided the earlier cascade.

Status distribution by directive

Xfail is shown as occurrences, with cross-platform distinct paths in parentheses. Stale combines stale xfail and stale not-applicable successes. Flake is failed/passed.

Directive	Observations	Distinct paths	Xfail	Not applicable	Stale	Flake	Resource	Unexpected failure	Ordinary pass
buildrun	2	1	0	0	0	0/0	0	0	2
buildrundir	6	4	0	2	0	0/0	0	4	0
compile	1,021	511	0	14	0	0/0	0	0	1,007
errorcheck	1,309	661	12 (6)	200	44	0/0	0	23	1,030
errorcheckandrundir	10	5	0	8	0	0/0	0	0	2
run	2,068	1,038	62 (35)	2	22	4/11	1	11	1,955
rundir	232	116	7 (4)	0	0	0/0	0	3	222
runindir	19	10	0	2	0	0/0	0	2	15
runoutput	44	22	2 (1)	0	1	0/1	1	2	37
Total	4,711	2,368	83 (46)	228	67	4/12	2	45	4,270

The remaining actual xfail distribution is therefore:

Family	Darwin	Linux	Both-platform occurrences	Distinct paths
compile	0	0	0	0
errorcheck	6	6	12	6
run	31	31	62	35
rundir	4	3	7	4
runindir	0	0	0	0
runoutput	1	1	2	1
Combined run*	36	35	71	40
All directives	42	41	83	46

The six actual errorcheck paths are bombad.go, fixedbugs/bug050.go, fixedbugs/issue13266.go, fixedbugs/issue20789.go, fixedbugs/issue4405.go, and fixedbugs/issue4776.go; all six fail on both targets.

Go-feature taxonomy of actual runnable xfails

Every actual run and run* xfail is assigned exactly once below. Each cell is distinct paths / platform occurrences.

Feature area	run	Combined run*
GC, finalizers, liveness, and memory profiling	12 / 22	12 / 22
Panic, recover, defer, nil, and bounds semantics	12 / 22	12 / 22
Stack, caller, traceback, and source-line metadata	6 / 11	7 / 12
Interface, reflection, type identity, and generics	3 / 5	6 / 11
Scheduler, goroutine, and channel execution	1 / 1	1 / 1
Init, link, codegen, and compiler lowering	0 / 0	1 / 2
Core language, numeric, and data semantics	1 / 1	1 / 1
Conservation total	35 / 62	40 / 71

The run* additions beyond direct run are cross-package interface/itab diagnostics, caller-frame metadata, and the generated rangegen.go codegen case.

This is the main-only baseline. I will post the rebuilt all-open-xfail-PR integration sweep separately, using the same matrix and status parser, so the delta is based on actual execution rather than YAML line counts.

[cpunion]

Go 1.26.5 all-open-xfail-PR integration sweep (Darwin + Linux)

The rebuilt integration sweep has completed against the same upstream main as the baseline.

• Upstream main: 41a336e (unchanged before and after this run)
• Integration head: 944b761
• Actions run: 31301568168
• Matrix: Go 1.26.5, darwin/arm64 + linux/amd64, directive-mode coverage, 4 shards per platform
• Included open xfail work: runtime,cl: add Go-compatible sampled memory profiling #2027, ssa,runtime: tighten recover to direct deferred calls #2033, ssa,runtime: make panic and recover Go-compatible #2034, cl,ssa,runtime: safely clear dead local finalizer roots #2036, cmd/compile: support type assertion diagnostics #2201, internal/packages: normalize Go compiler syntax diagnostics #2219, cmd/compile: validate noescape function bodies #2220, test/goroot: match more parser recovery diagnostics #2222, test/goroot: match EOF parser recovery diagnostics #2223, test/goroot: match malformed package clause diagnostics #2224, cl: tolerate invalid receivers during syntax preload #2226, test/goroot: match malformed selector recovery #2227, cl: preserve generic local type identity in nested closures #2229, test/goroot: match vareq recovery diagnostics #2230, cl: preserve zero values for blank struct fields #2231, runtime: preserve type metadata in assertion panics #2252, ssa: saturate float-to-uint32 conversions #2253, Superseded by #2316: preserve nil checks for unused dereferences #2256, ssa: preserve package identity for unexported methods #2258, and the expectation-only test/goroot: synchronize native run expectations #2264 refresh

All eight jobs completed full selected-case inventories. The run took 48m35s, essentially the same as the 49m03s main baseline. Coverage mode deliberately reports failure for stale expectations and unexpected failures.

Complete observed-status comparison

The Darwin inventory is 2,345 and the Linux inventory is 2,366, totaling exactly 4,711 selected observations and 2,368 cross-platform case paths.

Observed status	Main	Integration	Delta	Darwin	Linux	Integration distinct paths
Ordinary pass	4,270	4,380	+110	2,190	2,190	2,201
Actual xfail	83	43	-40	20	23	24
Not applicable	228	224	-4	114	110	114
Stale xfail: case actually succeeded	65	3	-62	1	2	2
Stale not-applicable: case actually succeeded	2	2	0	1	1	1
Known flake passed	12	10	-2	4	6	9
Known flake failed	4	9	+5	4	5	7
Explicit hosted-runner skip	0	4	+4	3	1	3
Resource guard stopped case	2	1	-1	0	1	1
Unexpected, unclassified failure	45	35	-10	8	27	28
Total	4,711	4,711	0	2,345	2,366	2,368 overall

The distinct-path column is status-specific and is not additive. The occurrence columns are mutually exclusive and conserve all 4,711 selected observations.

The four explicit hosted-runner skips are:

• chan/goroutines.go on Darwin: can exhaust the macOS runner thread limit
• fixedbugs/issue16016.go on Darwin: can exhaust the macOS runner thread limit
• fixedbugs/bug449.go on both targets: generated build exceeds the hosted-runner resource budget

The sole resource-guard stop is still fixedbugs/issue20780b.go on Linux. bug449.go is now skipped before consuming the runner instead of reaching the resource guard.

Status distribution by directive

Xfail is occurrences, with cross-platform distinct paths in parentheses. Stale combines stale xfail and stale not-applicable successes. Flake is failed/passed.

Directive	Observations	Distinct paths	Xfail	Not applicable	Stale	Flake	Host skip	Resource	Unexpected failure	Ordinary pass
buildrun	2	1	0	0	0	0/0	0	0	0	2
buildrundir	6	4	0	2	0	0/0	0	0	4	0
compile	1,021	511	0	14	0	0/0	0	0	0	1,007
errorcheck	1,309	661	4 (2)	196	2	0/0	0	0	19	1,088
errorcheckandrundir	10	5	0	8	0	0/0	0	0	0	2
run	2,068	1,038	35 (19)	2	3	9/9	2	1	8	1,999
rundir	232	116	1 (1)	0	0	0/0	0	0	2	229
runindir	19	10	0	2	0	0/0	0	0	2	15
runoutput	44	22	3 (2)	0	0	0/1	2	0	0	38
Total	4,711	2,368	43 (24)	224	5	9/10	4	1	35	4,380

The remaining actual xfail distribution is:

Family	Darwin	Linux	Both-platform occurrences	Distinct paths	Delta vs main occurrences/paths
compile	0	0	0	0	0 / 0
errorcheck	2	2	4	2	-8 / -4
run	16	19	35	19	-27 / -16
rundir	1	0	1	1	-6 / -3
runindir	0	0	0	0	0 / 0
runoutput	1	2	3	2	+1 / +1
Combined run*	18	21	39	22	-32 / -18
All directives	20	23	43	24	-40 / -22

The two remaining errorcheck paths are bombad.go and fixedbugs/issue4405.go; both fail on both targets.

Go-feature taxonomy of actual runnable xfails

Every actual run and run* xfail is assigned exactly once. Each cell is distinct paths / platform occurrences.

Feature area	run	Combined run*
GC, finalizers, liveness, and memory profiling	10 / 19	10 / 19
Panic, recover, defer, nil, and bounds semantics	2 / 3	2 / 3
Stack, caller, traceback, and source-line metadata	6 / 12	7 / 13
Interface, reflection, type identity, and generics	0 / 0	0 / 0
Scheduler, goroutine, and channel execution	1 / 1	1 / 1
Init, link, codegen, and compiler lowering	0 / 0	1 / 2
Core language, numeric, and data semantics	0 / 0	1 / 1
Conservation total	19 / 35	22 / 39

The run* additions beyond direct run are fixedbugs/issue29919.go (caller/frame metadata), rangegen.go (range-over-func generated code), and Linux-only convinline.go (numeric conversion semantics).

Important state transitions and remaining integration issues

• 38 main actual-xfail occurrences (21 paths) became ordinary passes. Another 2 occurrences, fixedbugs/issue52072.go on both targets, execute successfully but still carry an xfail, so coverage mode correctly reports them as stale.
• The expectation-only test/goroot: synchronize native run expectations #2264 refresh clears all 67 stale-expectation occurrences observed on main: 64 old stale xfails become ordinary passes, 2 stale not-applicable observations become passes, and the remaining old stale xfail is retained as a flake because it is not deterministic.
• The five stale observations in the integration result are newly exposed by stacking fixes: xfail entries for fixedbugs/issue52072.go (both targets) and inline_callers.go (Linux), plus the not-applicable entry for fixedbugs/issue18331.go (both targets). They are not stale on standalone main, so removing them from test/goroot: synchronize native run expectations #2264 would make that expectation-only PR depend on implementation PRs.
• recover1.go still fails on both targets, but the integration expectation commit removed its xfail. Those are the 2 main-xfail -> integration-unclassified-failure transitions and should not be counted as solved.
• Three observations changed from main pass to integration failure: fixedbugs/issue8047b.go on both targets and fixedbugs/issue27695.go on Darwin. Conversely, nine previously unclassified-failure observations became passes. The net unclassified-failure count is therefore down from 45 to 35, but the three pass -> fail changes still need individual attribution before merging the relevant PRs.
• Five known-flake observations changed from pass to fail in this run. They remain classified as flakes and explain why the raw flake-failure count is higher without implying five deterministic regressions.

No PR was merged as part of this integration run.