Reduce test-suite CI disk usage with profile.ci and runner cleanup

[cderv]

The test-suite workflow on ubuntu-latest hit "No space left on device" during cargo nextest run, even with the existing free-disk-space step already reclaiming 18.8 GB. The workspace has grown past PR #55's mitigation.

Root Cause

Two compounding pressures:

1. The default [profile.dev] emits full debuginfo (debug = 2), which roughly doubles target/ size on a workspace this big. Dependencies dominate target/ and carry no useful debug info for triaging q2 failures.

2. The workflow had both cargo build and cargo nextest run steps. Library crates compiled with --cfg test have a different fingerprint than dev builds and produce separate .rlib artifacts in target/<profile>/. With ~35 crates that runs into multi-GB duplication at peak disk.

Fix

Four independent levers, applied in one workflow change:

• Add [profile.ci] to the workspace Cargo.toml (inherits from dev, debug = "line-tables-only" on the workspace, debug = false on all dependencies). Local builds keep full debuginfo via the unchanged dev profile.
• Replace cargo build + cargo nextest run with a single cargo nextest run --tests --cargo-profile ci. Per the nextest design docs, nextest delegates compilation to cargo test --no-run, which makes the separate cargo build step redundant: the lib + bins + integration-tests that --tests selects strictly supersede cargo build's default targets (lib + bins, non-test). --tests rather than --all-targets because quarto-yaml's harness = false benches print prose reports that nextest can't enumerate as tests; the prior workflow never built benches anyway, so coverage matches. RUSTFLAGS=-D warnings continues to apply (it's a rustc env var, picked up regardless of which cargo subcommand drives the build).
• Enable remove_tool_cache: true on the existing endersonmenezes/free-disk-space step (~6 GB more). Verified safe: pre-installed rustup data lives in /etc/skel/.{rustup,cargo} and dtolnay/rust-toolchain installs to ~/.rustup/toolchains/, neither of which is in /opt/hostedtoolcache/.
• Add docker image prune --all --force + docker builder prune -a --force (~3-8 GB more). Pre-pulled docker images aren't used by this job.

Full rationale, measured numbers, and held-in-reserve options (swap removal, df -h diagnostic, larger runner) are in claude-notes/2026-04-28-ci-disk-space-and-profile-ci.md.

[cderv]

Follow-up fix: quarto-lsp integration test was hardcoding target/debug/q2, which broke under --cargo-profile ci (binary now lands at target/ci/q2).

Approach chosen: derive the profile directory from std::env::current_exe() — the test binary itself sits at target/<profile>/deps/, so backing up two parents lands in the same target/<profile>/ where the production bin is built. Profile-correct on all platforms with no env coupling.

Why not CARGO_BIN_EXE_<name>: that env var is package-scoped — Cargo only sets it for integration tests in the same package as the bin. Our LSP test lives in quarto-lsp while the bin is in quarto, so it's never set here.

Why not assert_cmd: assert_cmd::cargo::cargo_bin("q2") does exactly the same thing internally and is the idiomatic crate-based answer. We deferred adding it because this test doesn't use any of assert_cmd's actual value — no .assert(), no stdout matchers, just JSON-RPC over stdio. The trigger to switch is the first test that wants cargo run -- style command-driving with output assertions; at that point cargo_bin() is the standard binary-discovery helper that comes with the crate.

Full rationale in claude-notes/2026-04-28-ci-disk-space-and-profile-ci.md (Change 3 section).