refactor(config): per-domain decomposition {sites, ci} — carve, migration, and site-grammar cleanup

[claude-spd1]

Lands the decomposition-config carve (Oli's refactor/decomposition-config-shape, rebased onto feature/jax tip) plus the deferred pragmatics (phases D/E) and two follow-on cleanups that finish the carve's own stated intent.

What's here (by commit)

1. refactor(config) (Oli + Opus, rebased): decomposition apparatus out of PDConfig into per-domain decomposition: {sites, ci}; tiled site specs (glu_transformer/simple_mlp + LayerSelection) with explicit kept for toys; SiteTree/ArchFamily resolution (param_decomp/site_tree.py) — config→sites→chunks is now parse-free; load_run isinstance guard fixed (CIFn Protocol matched every impl) → ChunkwiseTransformerCIFn.
2. migrate(config): all 40 live repo yamls migrated by an in-repo tool (param_decomp_lab/migrations/decomposition_apparatus.py — comment-preserving line surgery, proven equal to its pure dict transform; each LM yaml's resolved site list verified byte-identical to the old canonical expansion). The 8 lab LM yamls already unparseable at tip (gpt2*, jose, llama-3.1-8b, pile/ss simple-mlp — pre-chunkwise ci_config schema) are left untouched.
3. test(config): tests + tooling ported off the old shape. Notable semantics: invalid site names are now parse-time rejections (Literal cs keys) instead of convert-time asserts; per-LAYER heterogeneous C is deliberately unrepresentable (tiling is what makes chunkwise chunks homogeneous by construction) — the old test_arbitrary_sites_with_per_site_c_convert is rewritten accordingly. Dead expand_wildcard_site_cs + wildcard pattern deleted. gen_probe_config output parses AND builds again (it had drifted pre-branch).
4. refactor(site-tree) ×2: finishing the branch doc's flagged costs on the typed-per-family fork — each target's KIND_ORDER now derives from the config Literal (get_args), the families live on their targets (llama8b.FAMILY / llama_simple_mlp.FAMILY), resolve_site_tree loses the dict relaunder + vocabulary-drift assert (unrepresentable now), and ArchFamily carries parse so the byte-identical canonical_site_cs/site_specs/tap_layer bodies live once in site_tree, family-parameterized (targets keep their two-arg public fns as one-line delegations — no caller churn).

Verification

• basedpyright clean workspace-wide; ruff clean.
• Equivalence goldens + SPEC.md untouched (pure config-shape change, no trajectory drift; git diff empty over tests/equivalence, tests/simple_mlp_equivalence, tests/stacked_parity, SPEC.md).
• Full --runslow -n4 suite: running at PR-open; will confirm on the task thread (targeted config/target/launch/finetune suites all green).
• experiments/invariance_check.py fails with the pre-existing fix(invariance_check): feed int tokens, not a float residual batch (SPEC D4) #940 float-tokens bug (bridge: fix-invariance-check-float-tokens) — reproduced identically pre-branch; not a regression of this PR.

Known deliberate gaps

• Stored run launch_config.yamls (152 dirs) don't parse post-carve — per Oli, explicitly out of scope for now ("not really worried about configs"); migrate_raw is written to serve that migration when wanted.
• Commits 1–3 were made with --no-verify (following the branch's own precedent when jax wasn't type-verifiable in the authoring env); later commits run the hooks. Everything is hook-clean at tip.

Crew-Address: task/targets-refactor
🤖 Generated with Claude Code

[claude-spd1]

Full-suite verification complete: 518 passed / 5 skipped / 11 xfailed under --runslow -n4; the 10 reported failures were all xdist worker crashes ("node down: Not properly terminated" — resource contention on a busy node), and all 10 pass on a serial rerun (10 passed in 141s). No test regressions. basedpyright + ruff clean; goldens + SPEC untouched.

Crew-Address: task/targets-refactor

[claude-spd1]

Rebased onto feature/jax tip past the Qwen3 ingestion (#987). What changed in the rebase:

• The carve's lab resolution now goes through feat(targets): ingest Qwen/Qwen3-8B-Base as a decomposition target #987's HF_MODEL_FAMILIES registry: any registered GLU model (Llama-3.1-8B, Qwen3-8B-Base) resolves its tiled glu_transformer c-spec against its own arch_config().n_layer — a Qwen3 run uses the same decomposition: schema out of the box.
• The vocabulary-derivation + family + site-grammar-sharing commits re-landed against targets/glu_transformer.py (their new home post-feat(targets): ingest Qwen/Qwen3-8B-Base as a decomposition target #987): glu_transformer.FAMILY / llama_simple_mlp.FAMILY, KIND_ORDER derived via get_args, shared canonical_site_cs/site_specs/tap_layer in site_tree.
• jose-ish.yaml (config(pile-4l): add jose-ish, a runnable migrated baseline #917, landed old-shape upstream) migrated by the in-repo tool.

basedpyright clean; all 41 live yamls parse (GLU ones build end-to-end); goldens (incl. the new qwen3_hf_parity) + SPEC untouched. Full --runslow suite running now — will confirm.

Crew-Address: task/targets-refactor

[claude-spd1]

Post-rebase full-suite verification: 525 passed / 6 skipped / 11 xfailed under --runslow -n4. The 8 reported failures: 7 were xdist worker crashes (node contention — all pass on serial rerun), 1 was an environment artifact of my shell (test_pretrained_target_converts_with_all_layers needs DATA_MOUNT, which my restarted session had lost; passes with it set). No regressions. CI is green on the pushed head as well.

Crew-Address: task/targets-refactor