fix(parser): eliminate 12h partition-phase hang on large models

[ebootheee]

What hung

A clean ete init on the full real models hung ~12h in the chunked emitter, right after [chunked] Partitioning 20 sheets.... That log line brackets partition_sheets() — which never returned. The next phase's "Emitting N sheet modules" log was never reached, so the stall was entirely inside partitioning (a misdiagnosis trap: "hung after Partitioning…" means the partition call itself, not the later par_iter sheet-module builder).

Root cause — a Round-2 regression

partition_sheets only needs sheet-level edges (which other sheets a formula references), but it called the range-expanding extract_refs. Commit 59810f6 (the cell-level dependency-graph feature) taught is_cell_ref to accept ranges — correct for that graph — but as a side effect extract_refs now expands every range to up to 1000 individual cell strings, pushes them, and de-dupes them. partition then throws away all the same-sheet ones. On the 1.62M-formula PP&E sheet (one rayon task = one core) that's ~10⁹ throwaway allocations → swap thrash → the 12h stall. The same trap sat one step downstream in detect_intra_sheet_cycles (would've been the next wall in the sheet-module phase).

Fix

• collect_sheet_deps() (new) — scans for Name! / 'Sheet Name'! tokens and records just the sheet name; no range expansion, no per-cell allocation. Tokenization mirrors extract_refs + the same cell-ref gate, so the detected dependency set is provably identical (parity test). partition_sheets now uses it.
• extract_refs_shallow() (new, top-left endpoint only) — used by detect_intra_sheet_cycles, restoring the pre-Round-2 behavior the known-good engines were built with (and avoiding spurious self-cycles from B10=SUM(B1:B10)).
• write_dependency_graph keeps the full expander — the dependency-graph.json / closures contract is unchanged. Added per-sheet progress to it (the heaviest remaining sequential step) so a slow-but-progressing pass no longer reads as a hang.

Behavior-preserving for the sheet-dependency set and engine output; ~2000× faster on range-heavy formulas (synthetic parity test: 2.66s → 1.3ms for 2,000 formulas in debug, gap widens under the memory pressure that caused the real hang).

Validation

• cargo test 17/17 (incl. parity + scale demonstration)
• npm run smoke 78/78 (engine accuracy, regenerated via the patched binary)
• test:depgraph 11/11 (full-expansion contract intact), test:runnable 20/20, test:engine 21/21 (cross-sheet cluster convergence intact)
• full npm test; clean build, zero warnings

Not addressed here (separate, still-open scaling walls → #22)

This unblocks the build. Making the result cheap to run as a calibration oracle is the next piece:

• engine.js eagerly imports ~800 MB of sheet modules (Node load-time wall)
• partition_sheets still clone()s every cell (peak-memory doubling)

Both fold into #22 (output-cone scoping / lazy sheet loading); ROADMAP/PLAN/HANDOFF updated to record them.

Note for re-running the regen: rebuild the release parser (cd pipelines/rust && cargo build --release) before ete init — the fix lives in the binary.

🤖 Generated with Claude Code