feat(pack): plan --fleet — prediction → placement (Act-2 opener) - #36

Summary

Act 2 opener — the highest-moat feature (audit judge 43.3). The contract's prediction becomes a placement decision. Because the predicted peak is computed from the index header alone (no run, no read I/O — milliseconds), is a conservative upper bound (sound since the Act-1 soundness fix recorded predicted_peak_rss_bytes), and is additive, a scheduler can sum predicted peaks across co-located jobs and prove a node fits before launching a byte.

No emergent-peak caller (GATK, DeepVariant) can do this: their peak is only known after a possible OOM-kill, so every co-location is a gamble. Copying it requires first making peak predictable — which is the entire Rosalind thesis.

What's new

• src/call/pack.rs — a pure, deterministic first-fit-decreasing bin-packer over predicted peaks. Packed(...) is a proof: every node's summed peak ≤ capacity. A job larger than a whole node, or a set that needs more than --nodes, is NoFit. Ties break by label → the schedule is reproducible (the contract's determinism extends to placement).
• rosalind pack --jobs <tsv> --node-mb N [--nodes M] — read each job's index header → predicted peak → pack onto nodes, or refuse (exit 3) with an actionable message. Human-readable or --json.
• rosalind plan --index --json — the predicted-peak breakdown as one line of JSON for a workflow engine to read programmatically.

$ rosalind pack --jobs jobs.tsv --node-mb 64
pack: 3 job(s) → 1 node(s) of 64 MiB — every node proven within capacity
  node 0: 48 / 64 MiB  [a.idx, a.idx, b.idx]

$ rosalind pack --jobs one-big.tsv --node-mb 16        # job > node
pack: REFUSE — job 'big.idx' predicted peak 19229088 B exceeds the node capacity 16777216 B …   # exit 3

Test plan

• 5 packer unit tests — incl. the proof property (every node ≤ capacity), determinism (order-independent), job-bigger-than-node refusal, --nodes cap
• tests/pack.rs integration — fit (proof line) / refuse (exit 3) / --json
• plan --index --json emits the machine-readable prediction
• Full suite green (31 sections); rustc 0 warnings; additions clippy-clean; smoke-tested end-to-end on real indexes

Why this, now

It's the single highest-leverage compounding move: it makes the contract a fleet property, and it's only sound because Act-1 PR-A made the predicted peak a true upper bound. Next Act-2 candidates (from the audit): ColumnKit SDK (46.3), bounded gVCF (42.7), verify-attest signed receipts (43).

🤖 Generated with Claude Code