ci: split platform crates from workspace catch-all

[ruvnet]

Summary

• move 17 service/distributed crates and 9 native Node bindings into a dedicated core-platform test shard
• reduce the core-and-rest catch-all from 78 packages to 50
• move ruvector-graph-condense-wasm into the WASM shard and correct the duplicated decompiler exclusion

Why

The post-merge and PR validations for #821 remained in the catch-all test step beyond the historical 155–165 minute completion window. Five preceding runs timed out at 240 minutes with compile/link processes still active. This additional split keeps full test/doctest coverage while reducing cold-cache linker pressure.

Validation

• workflow YAML parses successfully
• all 26 core-platform package names exist in workspace metadata
• no core-platform package overlaps another explicit shard
• catch-all package count verified at 50
• git diff --check passes

[ruvnet]

Review — 2 blocking issues found and fixed

Reviewed against this PR's own CI run logs. The shard split arithmetic is correct, but two defects made it ineffective. Both are fixed in 09cccfe.

---

Finding 1 (critical): every --exclude in the catch-all shard was silently discarded

packages: is a YAML folded scalar (>-), and the core-and-rest entry had #-prefixed lines inside the scalar. Those are content, not YAML comments. Folding joins the block onto one line, and when the run: step expands ${{ matrix.packages }} the shell treats the first # as a comment and truncates everything after it.

The command that actually ran was:

cargo nextest run --no-fail-fast --workspace

0 of 162 --exclude flags survived. The catch-all shard was building and testing all 210 workspace crates on every run. That explains the behaviour the timeout bumps in iters 228-240 were chasing: every shard split landed so far has been inert, so the catch-all kept drifting into the cap no matter how many crates were hoisted out of it. The doctest step (cargo test --doc ${{ matrix.packages }}) had the same truncation.

core-and-rest-wasm had the same defect, but its inline comment was trailing prose with no flags after it, so all 29 -p flags survived. Fixed anyway so the pattern is not copied forward.

Fix: moved every comment above the packages: key at mapping level, where YAML strips it — the pattern the new core-platform entry already used correctly. The folded scalars now contain only --workspace, --exclude <crate>, and -p <crate> tokens.

Finding 2 (critical): a hung test can burn the entire job budget

ruvector-delta-index::tests::test_insert_and_search (crates/ruvector-delta-index/src/lib.rs:731) hangs indefinitely in DeltaHnsw insert/search rather than failing. It sat in a nextest SLOW loop for 3h52m on run 31675583441 until core-platform was cancelled at the 240-minute cap. The test builds its input from an unseeded rand::thread_rng(), so the hang is nondeterministic — which is why it only surfaced once this PR gave the crate its own shard.

There was no .config/nextest.toml, so nothing bounded it.

Fix:

• Added .config/nextest.toml with slow-timeout = { period = "120s", terminate-after = 5 } — any test is killed and reported as a normal failure after 10 minutes instead of running forever.
• Held ruvector-delta-index out of CI entirely (removed from core-platform, kept in the catch-all --exclude list) pending ruvector-delta-index test_insert_and_search hangs indefinitely (DeltaHnsw insert/search) #825.

---

Shard coverage — verified

Parsed the workflow with PyYAML and diffed each shard's effective package set against cargo metadata:

shard	mode	excludes	effective pkgs
vector-index	-p list	0	6
rvagent	-p list	0	10
ruvix	-p list	0	16
ml-research-heavy	-p list	0	4
ml-research-rest	-p list	0	6
core-and-rest-heavy	-p list	0	9
core-and-rest-wasm	-p list	0	29
research-nightly	-p list	0	12
core-and-rest-examples	-p list	0	36
core-platform (new)	-p list	0	25
core-and-rest	--workspace	162	50

• 26 crates moved into core-platform by this PR; 25 remain after holding out ruvector-delta-index. Each has a matching --exclude in the catch-all.
• No crate is built twice — the previously double-built ruvector-decompiler is fixed (it is in core-and-rest-heavy and excluded from the catch-all exactly once).
• 203 of 210 crates covered. The 7 uncovered are all deliberate: the 4 hailo crates (hailort-sys, ruvector-hailo, ruvector-hailo-cluster, ruvector-mmwave), ruvllm-wasm (wasm-target-only tests SIGABRT on native), timesfm (tested by its own dedicated step with --features candle), and ruvector-delta-index (ruvector-delta-index test_insert_and_search hangs indefinitely (DeltaHnsw insert/search) #825).

Only after Finding 1 is fixed do those 162 excludes take effect — so this is the first run where the shard split will actually reduce the catch-all's workload, from 210 crates to 50.

Note on scope

Two pre-existing --exclude targets are not workspace members: ruvector-postgres and timesfm-wasm. Harmless and unrelated to this PR, so left alone.

---

Review and fixes generated by Claude.