feat(cache): guest-side prune, asserted against an independent oracle

Measured first, as thread entry #12 requires: a cold furyrs pipeline is 134.7s,
target/ is 3.0 GiB raw and 2.5 GiB pruned, and the stored entry is 645 MiB. A
warm job pays 8.1s (upload + extract) to skip ~120s of dependency compilation.
Publishing costs 15.7s but only on a miss. Build it.

The assertion in the plan turned out to be worthless and is replaced. It
re-scanned for */.fingerprint/<member>-* after deleting exactly those paths, so
asked to catch a planted fury-core fingerprint it exited 0 -- the prune had
removed the evidence before the check looked. An assertion sharing its globs
with the deletion can only prove that rm ran; it cannot see the failure that
matters, which is a glob that never matched the artifact in the first place.

The replacement asks a different oracle. Cargo records in each .d depfile the
sources an artifact was built from: dependencies point into the registry,
anything built from workspace source points back into the workspace. After a
correct prune, zero surviving depfiles reference workspace source outside
target/. Verified negatively too -- an artifact named totallyunrelated-abc123.d
that no prune glob matches, declaring a dependency on fury-core/src/lib.rs,
makes the script exit 2 and refuse to publish.

zstd is already present on the template, so no rebuild is needed; the gzip
fallback stays for the plain LinuxBuild2404.

plans/cargo-dependency-cache-measurements.md

@@ -0,0 +1,93 @@
+# Cargo dependency cache — measurements
+
+Measured 2026-07-22 on a `LinuxRust2404` linked clone (4 vCPU, 3.9 GB RAM,
+`rustc 1.97.1`, `zstd` present at `/usr/bin/zstd`), running furyrs `b8ea567`
+with the `build-command` block from its `.gitea/workflows/ci.yml` verbatim.
+
+Entry #12 of the furyrs thread set the rule: measure before building. There it
+killed the registry cache — 9.3s of saving for +800 MB on the template. Here
+the numbers point the other way, decisively.
+
+## Baseline
+
+| | |
+|---|---|
+| Cold pipeline, empty `target/` | **134.7s** |
+
+Matches the 133s furyrs reported for job 1021 and the 126.6s measured in #12,
+so the clone is representative.
+
+## Sizes
+
+| | Bytes | Human |
+|---|---:|---:|
+| `target/` raw | 3,206,670,175 | 3.0 GiB |
+| `tar` of raw `target/` | 3,215,503,360 | 3.0 GiB |
+| `tar` + `zstd -3 -T0`, raw | 810,108,028 | 773 MiB |
+| **`target/` after prune** | 2,719,971,105 | **2.5 GiB** |
+| **entry as stored** (`tar --zstd`, pruned) | 676,585,779 | **645 MiB** |
+
+Pruning strips 487 MB (15%) of raw `target/`, which is the workspace crates'
+own artifacts — 42 fingerprint directories across the six members
+(`fury-core` 11, `fury-gui` 12, `fury-ipc` 6, `fury-backend` 5, `fury-cli` 5,
+`furyd` 3). 1,928 dependency artifacts and 903 dependency fingerprints survive.
+
+## Costs
+
+| Step | Seconds | Paid by |
+|---|---:|---|
+| prune + assert | 0.21 | publish |
+| pack (`tar --zstd`) | 10.88 | publish |
+| fetch guest→host (140 MiB/s) | 4.62 | publish |
+| **publish total** | **15.7** | **only on a cache miss** |
+| upload host→guest (179 MiB/s) | 3.60 | restore |
+| extract | 4.52 | restore |
+| **restore total** | **8.1** | **every warm job** |
+
+## Verdict: build it
+
+A warm job pays **8.1s** to skip roughly **120s** of dependency compilation —
+the ~490 external crates that a committed `Cargo.lock` keeps bit-identical
+between runs. Net saving ≈ 112s on a 134.7s job, about **83%**.
+
+The publish path costs 15.7s but runs only on a miss, i.e. once per
+`Cargo.lock` change, so it is amortised across every run in between.
+
+Transfer is not the bottleneck anyone feared: the virtual NIC moves the entry
+at 140–180 MiB/s, so 645 MiB crosses in under 4 seconds each way.
+
+## Notes that changed the implementation
+
+**`zstd` is already in the template.** `/usr/bin/zstd` is present on
+`LinuxRust2404`, so the gzip fallback is dead code in practice and no template
+rebuild is needed. Keep the fallback anyway — it costs one branch and keeps the
+feature working on a plain `LinuxBuild2404`.
+
+**The name-based assertion was worthless and was replaced.** The plan's
+original check re-scanned for `*/.fingerprint/<member>-*` after deleting
+exactly those paths. Asked to catch a planted `fury-core-deadbeefcafe`
+fingerprint, it exited **0** — the prune had deleted the evidence before the
+assertion looked. An assertion built from the same globs as the deletion can
+only prove that `rm` ran; it is blind to the failure that matters, a glob that
+never matched the artifact at all.
+
+The replacement asks a different oracle. Cargo records in each `.d` depfile the
+source files an artifact was built *from*: a dependency points into
+`/opt/rust/cargo/registry/...`, anything built from workspace source points
+back into the workspace. After a correct prune, **zero** surviving depfiles
+reference workspace source outside `target/`. Verified negatively as well — an
+artifact named `totallyunrelated-abc123.d` (which no prune glob matches)
+declaring a dependency on `fury-core/src/lib.rs` makes the script exit **2**
+and refuse to publish.
+
+| Check | Clean tree | Planted fingerprint | Planted foreign artifact |
+|---|---|---|---|
+| name-based (original) | 0 | **0 — missed it** | 0 — missed it |
+| depfile-based (shipped) | 0 | 0 (prune removed it, correctly) | **2 — refused** |
+
+**Feature variants coexist in one entry, as predicted.** `deps/` holds
+`libclap-943062c20e623534.rmeta` and `libclap-abe395d6778d44ce.rmeta` side by
+side — the same crate compiled under two different feature resolutions. This is
+why the feature set is not part of the cache key: cargo already separates the
+variants by hash, so one entry serves all of furyrs's four feature
+configurations instead of fragmenting into four.