Drop third-party dependencies from the default build (v1.1.0)

[jamesgober]

Do not merge, tag or publish yet: owner decisions needed (see below).

Summary

With default features, error-forge pulled 8 third-party crates into every dependent's normal build, while the dependents (throttle-net, rate-net, better-bucket and the rest) use only the ForgeError trait. Both always-on dependencies are now gone. The default graph is error-forge plus its own error-forge-derive, which has no dependencies of its own. No public item was added to, removed from or changed in error-forge, and default = [] is unchanged.

What needed it (1.0.2)	Crates	Replacement
define_errors! constructor names (fn [<$variant:lower>])	pastey	~130-line proc macro in error-forge-derive on proc_macro alone (error_forge::__private::paste!, internal). syn/quote/proc-macro2 move behind a new default-on derive feature of the derive crate; error-forge depends on it with default features off
recovery::CircuitBreaker internal state	parking_lot, lock_api, parking_lot_core, scopeguard, smallvec, cfg-if, windows-link/libc	std::sync::Mutex via PoisonError::into_inner, so it still never panics. The success path now takes the lock once instead of twice

All other third-party deps (serde, log, tracing, async-trait, rand, and syn for derive) were already behind opt-in features and are unchanged.

Evidence

cargo tree -p error-forge -e normal, default features, excluding self

Target	v1.0.2	this branch
Windows x86_64	8 third-party	1 in-house (error-forge-derive, 0 deps)
Linux x86_64	8 third-party	1 in-house
macOS aarch64	8 third-party	1 in-house
all features (Windows)	36	29

Clean debug build of a dependent (Windows, CARGO_INCREMENTAL=0, median of 5): -j32 1.09 s → 0.78 s; -j1 2.67 s → 0.99 s.

CircuitBreaker::execute (release, no-op closure, ns per call):

	Win	Win, 8 threads	Linux	Linux, 8 threads
1.0.2 (parking_lot, 2 locks)	41.2	346	38.9	136
branch with parking_lot (1 lock)	33.8	260	30.6	101
this branch (std, 1 lock)	34.2	182	31.3	122

Uncontended, std and parking_lot are equal. Under heavy synthetic contention parking_lot is about 20 ns/call better on Linux and about 80 ns/call worse on Windows. Next to the I/O a breaker guards, neither difference matters, so it doesn't meet the "necessary and measurably better" bar. Compared with 1.0.2, the branch is faster everywhere.

Gates

• cargo fmt --all -- --check
• clippy -D warnings on 1.95 and 1.99: all-features, default, no-default, derive crate alone
• Tests: all 9 CI matrix combos, plus error-forge-derive, async log tracing jitter, and all-features (140, incl. trybuild UI) on Windows 1.95. Linux 1.99 all-features: 140
• MSRV 1.81.0: build plus 139 tests (UI skipped, as in CI); minimal-versions emulation builds on 1.81
• RUSTDOCFLAGS=-D warnings cargo doc
• cargo audit --deny warnings clean; cargo deny check ok (same warnings as main)
• cargo semver-checks -p error-forge --baseline-version 1.0.2: 196 pass, no semver update required. The tool can't check proc-macro crates, so error-forge-derive was reviewed by hand: it only adds a default-on derive feature and a hidden macro
• cargo publish --workspace --dry-run: both crates package and verify
• New CI gate: the default dependency graph must contain no third-party crate (--target all)

New tests: tests/paste_test.rs covers constructor naming (acronyms, r# variants, non-ASCII, digits) and the macro's modifiers and nesting. tests/ui/fail/paste_bad_modifier.rs covers each error path. A Type variant gives the same expected identifier, found keyword error as 1.0.2 (checked against the published crate).

Decisions needed from the owner

1. STABILITY.md contract. The crate's STABILITY.md (binding per DIRECTIVES D-5) said "Removing any of the always-on deps requires a 2.0.0 bump". It also listed "adding a runtime dependency that is not gated behind an opt-in feature" as breaking; error-forge-derive goes from optional to always-on. VERSIONING.md, which is authoritative across the portfolio, classes neither change as breaking: the "would user code still compile?" test passes, and semver-checks agrees. The 1.0.2 removal of thiserror set the same precedent. This PR narrows those clauses to third-party dependencies and public-API exposure, with a paragraph explaining why. Please confirm this is MINOR rather than MAJOR.
2. Version and schedule. The PR bumps to 1.1.0: the derive crate gains a feature, and the perf gains are worth calling out. The 2.0 plan already reserves 1.1.0 for the runway (R1–R10). Either ship this as 1.1.0 and move the runway to 1.2.0, or hold this PR and fold the runway items in before tagging.
3. The open Dependabot PR (cargo-minor-and-patch) also touches Cargo.lock, so whichever merges second gets a trivial lockfile conflict.

After approval, the release steps are: date the CHANGELOG section, tag v1.1.0, publish error-forge-derive and then error-forge. Local notes are in .dev/audits/2026-10-14.md and .dev/release/v1.1.0.md (gitignored).