ci(test): reclaim ubuntu runner disk + drop test debug-info

[coseto6125]

Why

PR #291 (T4-7 MirrorsField) hit a misleading CI failure on ubuntu-latest
(run 26240379100):

error: failed to build archive at `.../libtantivy-895456c1b8c7e622.rlib`:
  No space left on device (os error 28)
error: could not compile `ecp-analyzer` (test "callmeta_c_cpp") due to 1 previous error

Root cause: runner disk exhaustion, not a code bug. macOS test passed,
clippy/audit passed. The "test callmeta_c_cpp" line is a cascading cargo
report — the actual linker failure was os error 28.

The math: 14 tree-sitter parsers × ~30 test binaries × default debug-info
(true) ≈ 30 GB in target/debug/deps/. Default ubuntu runner free disk
is ~14 GB. Inevitable as the workspace grew.

What

Two stacked mitigations, both ubuntu-targeted but the env var applies to
all OSes since it's harmless elsewhere:

1. jlumbroso/free-disk-space step before checkout (ubuntu-only)
   reclaims ~30 GB by dropping tool-cache + android SDK + dotnet + haskell.
   large-packages left off — adds ~5 min for ~1 GB. swap-storage off —
   risky if any step needs swap. Action pinned to v1.3.1 SHA per repo
   convention (matches step-security/harden-runner, actions/checkout).
2. Job-level CARGO_PROFILE_TEST_DEBUG=0 drops DWARF symbols from
   test binaries (~70 % size reduction). Assertions and overflow checks
   stay on; only stack-trace symbolication is lost — cargo nextest
   output already includes test names + file:line via panic hooks, so
   diagnostic value is preserved.

Combined: ~14 GB baseline + ~30 GB reclaimed − ~9 GB build ≈ ~35 GB
buffer. Future language / dependency growth has headroom.

Test plan

• Lint jobs pass (actionlint validates the new step)
• All three matrix OSes pass Run tests (nextest)
• After merge: re-run feat(graph): SchemaField nodes + MirrorsField heuristic edges (T4-7) #291's failed Test on ubuntu-latest job →
  expect green

Notes

• Does not touch ci: short-circuit heavy jobs for docs-only PRs (no branch-protection deadlock) #287 (docs-skip short-circuit) — orthogonal fix; this PR
  reduces resource pressure when the test job does run.
• Does not touch any in-flight feat PR. Pure CI plumbing.