fix(build): stop two builds fighting over target/<profile>/deps/liboxidex.*

[swackhamer]

Two builds were fighting over target/<profile>/deps/liboxidex.*. Both are the
same bug class: [lib] crate-type = ["lib", "staticlib", "cdylib"] makes cargo
drop -C extra-filename, so every configuration of the lib shares one set of
filenames and whichever writer finishes last wins.

1. The C FFI test's nested build (finishes #639)

#639 stopped tests/ffi_c_integration.rs's nested cargo build --lib from
building the default feature graph over the rlib the doctests link — the
E0432 on both parsers::magika_detector doctests. It forwarded the feature
graph and recorded that the nested build now "writes nothing at all".

Measured on cargo test --all-features -v, that part is false. A nested cargo
inherits no profile either: it ran under dev while cargo test builds the
lib under test, which are different units here ([profile.test]
opt-level = 2, codegen-units = 4 vs [profile.dev]'s 0/16, read off the
rustc invocations). 44s into the FFI test it logged Compiling oxidex /
Finished `dev` profile, and target/debug/deps/liboxidex.rlib went
119,540,352 → 99,273,872 bytes underneath the doctests.

A test can't learn the profile name cargo invoked it under, so a nested build
can't be made to match one. It doesn't need to build at all: cargo builds the
lib as a dependency of every integration-test target, all three crate types in
one rustc call. The test now derives its link directory from
env::current_exe() and links what's already there.

It has to be target/<profile>/deps, not the profile root — cargo uplifts a lib
target only when it's a requested target. After
cargo clean -p oxidex && cargo test --all-features --no-run, deps/ holds all
three artifacts and target/debug/ holds none; the nested build was doing that
uplift, which is the only reason the old -L target/debug resolved.

2. cargo test --release races two lib units

Any --release invocation whose graph also contains test targets compiles
src/lib.rs twice: unwind (cargo forces it on test targets and their deps) and
abort (the three bins are final artifacts reading panic from
[profile.release]). Measured with
cargo test --release --all-features --features tag-comparison-binary --test ffi_c_integration --no-run -v:
two --crate-type lib invocations identical but for -C panic=abort, neither
carrying -C extra-filename, 4 output filename collision warnings
(rust-lang/cargo#6313), then a race. When abort wins, dependents fail with
"requires panic strategy abort … incompatible with … unwind" on unchanged
source. This is what the [profile.bench] comment already documents, with
[profile.release] as the abort source.

That cure — pin panic = "unwind" — doesn't transfer: [profile.release]'s
panic = 'abort' is what shipped binaries want. So the override is scoped to
the invocations that need it (unwind in the justfile). The bins those recipes
build are throwaway test-run artifacts; just build-bin-release,
cargo build --release, the tag-comparison bin recipes and the Release Build CI
job carry no prefix, so released binaries still abort. With the prefix the
same command reports 0 warnings and a single lib unit.

CI is unaffected either way — it runs no --release test step, which is why
this never showed as a red build, only as an intermittent local just test.

Not fixed here

The root enabler is the dropped -C extra-filename. The only complete cure is
moving the C ABI into its own crate so oxidex can return to
crate-type = ["lib"]; that touches cbindgen, the committed header, the python
bindings and the packaging assets, and CI can't validate it. Wants its own change.

Verification

• cargo test --all-features (post-merge with origin/main): exit 0, 0 failures,
  doctests 221 passed / 0 failed. On a warm re-run the SHA-256 of
  target/debug/deps/liboxidex.rlib is identical before and after — nothing
  overwrites it now.
• FFI suite 45.44s → 0.69s; the 44.75s was the nested rebuild.
• cargo test --release --all-features --features tag-comparison-binary --test ffi_c_integration:
  exit 0, 41 C assertions pass, now linking the release cdylib that invocation
  actually built.
• cargo fmt --check clean; cargo clippy --all-features --test ffi_c_integration -- -D warnings exit 0.
• just --dry-run test shows the prefix expanded; shipping recipes verified unprefixed.

🤖 Generated with Claude Code