meta: stop caching Rust toolchains in wild CI (setup-soldr seal-at-install + no-scan off mode, soldr target self-heal)

[zackees]

Goal

Stop caching Rust toolchains in wild's CI, and do it without the scan cost setup-soldr pays today.

• soldr remembers the toolchain it installed (its prepare memo) and heals it when it disappears.
• setup-soldr gains a toolchain-cache-off mode with no scans. Where the cache is used, it's saved once, right after install.
• wild adopts the new @v0 with toolchain caching off.

Validation gate: this meta-issue is done only when

wild (zackees/wild) no longer has any Rust toolchain cached, verified on a main push and a PR run with the new zackees/setup-soldr@v0:

• gh cache list -R zackees/wild --key solo-toolchain is empty;
• no setup-soldr job reports snapshot_* or solo_restore in setup phase totals;
• each container job's toolchain phase is ≤ 12s;
• CI is green, including every cross target and wasm.

The full gate (G1–G5) is in #18.

Why

Measured in this repo (details and run IDs are in each sub-issue):

• 2.1 GB of cache: 14 solo-toolchain-* entries hold identical rustc bytes duplicated per soldr version and branch, putting the repo cache at 10.2 GB, over the 10 GB limit.
• 25–30s of scanning per container job. setup-soldr walks the image's 135k-file /usr/local/rustup up to three times, even on cache hits.
• Corruption: the toolchain cache is saved from a file list fixed during setup but read at job end, which caused the PPC64LE E0463 in ci: adopt soldr locally and setup-soldr in CI workflows #16.

Sub-issues, in order

1. fix(solo-toolchain-cache): seal the toolchain at install time with no scans and a soldr-version-free key setup-soldr#525: seal the toolchain at install time with no scans. Includes the addendum: with the cache off, no scans and nothing written (T8/E7), then cut a release and move v0.
2. solo-toolchain-cache saves a pre-build file list with post-build contents: a target installed mid-build poisons the shared tnone key (E0463) setup-soldr#507: the existing corruption bug (post-build contents saved under a pre-build file list). Closed by Write unreferenced internal symbols for non-empty sections wild-linker/wild#525.
3. fix(toolchain): self-heal missing target std libraries; readiness and memo must check rust-std files, not only rustup metadata soldr#3376: self-heal target libraries that are claimed but missing. Readiness, the memo and the ensure smoke check verify the rust-std files, not only rustup's metadata.
4. ci: adopt new setup-soldr v0 with toolchain caching off; gate on zero cached toolchains #18: wild adopts the new v0 with solo-toolchain-cache: false everywhere, purges the old entries, and runs the validation gate.

Steps 1 and 3 are independent and can run in parallel. Step 4 needs step 1 to be released.

Decisions

• wild pays the rustup install (about 7–9s per job) instead of caching toolchains, removing the cache space, the scans and the poisoning bug.
• setup-soldr keeps a correct, opt-in toolchain cache for other consumers. wild turns it off explicitly, so the gate doesn't depend on setup-soldr's default.
• wild stays on the floating @v0, so the fix lands through the setup-soldr release.

[zackees]

#18 validation gate results

setup-soldr v0 was promoted a71cd1c9 → dfbe9627. That head includes zackees/setup-soldr#530, which skips toolchain snapshots when solo-toolchain-cache is off. Contract: zackees/setup-soldr run 36220387677. Canary: FastLED/fbuild run 36224341140.

wild#20 (merged) sets solo-toolchain-cache: false on every setup-soldr step. All 14 solo-toolchain-* entries were purged before the runs.

Runs on the new v0:

• main: workflow_dispatch run 36225933912: success
• PR: ci: #18 gate validation run (do not merge) #21 (validation-only, closed): CI run 36225934628 and performance guard 36225934689, both success

Toolchain phase (setup phase totals)

Job	Before (run 36192705272)	main after	PR after
Build Linux (RISC-V)	19.5s (scans 10.6s)	9.3s {rustup_install=9.3s}	9.9s {rustup_install=9.9s}
Build Linux (PPC64LE)	21.6s (scans 12.3s)	11.7s {rustup_install=11.7s}	7.7s {rustup_install=7.7s}
Build Linux (LoongArch64)	25.1s (scans 16.5s)	9.3s {rustup_install=9.3s}	14.3s {rustup_install=14.3s}
Build wasip1	12.4s (scans 5.0s)	10.3s {rustup_install=10.3s}	10.4s {rustup_install=10.4s}
Test Wasm	15.0s (scans 6.2s)	8.4s {rustup_install=8.4s}	7.3s {rustup_install=7.3s}
minimal-versions	23.2s (scans 16.7s)	11.2s {rustup_install=11.2s}	10.6s {rustup_install=10.6s}
Test Mach-O (AArch64)	–	6.5s {rustup_install=6.5s}	7.2s {rustup_install=7.2s}
Build Windows (x86_64)	–	10.1s {rustup_install=10.1s}	12.3s {rustup_install=12.3s}
Linux performance guard	–	(PR-only)	1.4s {rustup_install=1.4s}

Example full lines, LoongArch64: main toolchain=9.3s {rustup_install=9.3s} ... total=25.2s; PR toolchain=14.3s {rustup_install=14.3s} ... total=54.0s.

Gate

• G1: no toolchain cached. gh cache list -R zackees/wild --key solo-toolchain returns 0 entries after both runs. The PR ci: turn off setup-soldr's toolchain cache in every job (#18) #20 run and the main push after the purge created none either.
• G2: no scans or toolchain-cache activity. None of the 17 setup-soldr jobs logs snapshot_*, solo_restore or solo-toolchain-cache: key=. Every job logs toolchain: solo-toolchain-cache off — skipping toolchain snapshots (#525). The toolchain phase equals rustup_install exactly.
• G3: container jobs ≤ 12s. Met on main: all six container jobs took 9.3–11.7s. The PR run missed on one job: LoongArch64 took 14.3s, all of it rustup_install, i.e. download time; the other five took 7.3–10.6s. No scan time is left to remove, so the remaining variance is rustup/network.
• G4: CI green, no E0463. Both runs are green, including every cross target and wasm. E0463 appears in no log.
• G5: Actions cache down ≥ 2 GB. Toolchain entries went from 14 / 2.25 GB to 0. Total usage still went from 10.68 GB (71 entries) to 11.19 GB (73 entries). The validation runs saved 22 new entries (3.87 GB): 16 from main runs and 6 from PR ci: turn off setup-soldr's toolchain cache in every job (#18) #20, which still ran on the old v0 that saved PR caches. The growth is in cook-base (4.38→4.99 GB), cook-delta (0.20→1.37 GB) and setup-soldr-buildcache (2.24→3.15 GB). PR ci: #18 gate validation run (do not merge) #21, on the new v0, saved nothing (setup-soldr#527). GitHub LRU-evicts above 10 GB.

gh cache list summary by key prefix

prefix	before	after
solo-toolchain	14 / 2.25 GB	0 / 0 GB
cook-base	9 / 4.38 GB	11 / 4.99 GB
setup-soldr-buildcache	11 / 2.24 GB	14 / 3.15 GB
cook-delta	5 / 0.20 GB	12 / 1.37 GB
setup-soldr-cargoregistry	16 / 0.53 GB	17 / 0.57 GB
other (v0-rust-*, soldr-mini)	16 / 1.08 GB	19 / 1.11 GB
total	71 / 10.68 GB	73 / 11.19 GB

Not done

• The exp/soldr-cache-toolchain probe branch is kept. ci: adopt new setup-soldr v0 with toolchain caching off; gate on zero cached toolchains #18 deletes it only once fix(solo-toolchain-cache): seal the toolchain at install time with no scans and a soldr-version-free key setup-soldr#525 has merged. Only Write unreferenced internal symbols for non-empty sections wild-linker/wild#525's T8/E7 cache-off addendum shipped (linker-diff needs to be taught about: rel.match_failed.R_AARCH64_TLSDESC_ADR_PAGE21 wild-linker/wild#530); its seal-at-install rework is still open.
• ci: canary setup-soldr dfbe962 before v0 promotion FastLED/fbuild#1471 (the canary) is still open. Besides the pin, it fixes ban_std_pathbuf/ban_raw_fbuild_path/ban_unrooted_tempdir dylint violations that feat(bench): per-phase cold-build timing breakdown in Blink benchmark (#1465) FastLED/fbuild#1470 left on fbuild main.

[zackees]

G5 update: removed 34 cache entries (3.85 GB) that no run can restore any more:

• 17 from merged PRs bench: compare fork revisions against upstream (#14) #17 and ci: turn off setup-soldr's toolchain cache in every job (#18) #20 (2.33 GB). PR-scoped caches are only restorable by the same PR's runs.
• 17 from the exp/soldr-cache-toolchain probe branch (1.52 GB). Its run logs, cited in fix(solo-toolchain-cache): seal the toolchain at install time with no scans and a soldr-version-free key setup-soldr#525, are untouched, and the branch is kept.

gh cache list -R zackees/wild now shows 39 entries, 7.33 GB, all on refs/heads/main, down 3.35 GB from 10.68 GB (2.87 GB below the 10.2 GB baseline). solo-toolchain entries: 0. G5 passes.

Only G3 remains partial: in the PR run, LoongArch64's rustup download took 14.3s; on main, all container jobs took ≤ 11.7s.