Derive a coverage grid before writing scenarios, and sample it deliberately

[KarthikAvinashFI]

What

Rewrites the write-scenarios skill so the scenarios stage derives a coverage grid from the agent's contract before writing anything, samples that grid deliberately, fans the writing out across the existing parallel writer sessions, and closes with a coverage report. Enables the parallel path in the hosted image and sizes its batches.

No code changes. The diff is one skill file and three ENV lines plus a revision bump in Dockerfile.hosted.

Why

Measured across 718 previously generated scenarios: 5 of the 12 canonical task operations were effectively untested (diagnose and navigate at 0%, compare at 1%), and only 5% carried any adversarial or safety overlay, with impersonation at zero. The suites collapse into create/execute because nothing tells the writing model what the space of situations is or which cells it must not skip.

The skill now instructs the model to:

• derive task intent as a closed 12-operation set crossed with the contract's own domain objects, pruned, so intent coverage is exhaustive by construction rather than imagined
• treat counterparty, disposition, channel, interaction and adversarial overlay as explicit axes, with one off-baseline axis per scenario so results stay attributable
• distinguish world-backed overlays (impersonation, fraud: need setup and proof) from prompt-side ones (injection, pressure: live in the instruction)
• mask incoherent cells, then sample with a hard-required checklist: emergency, injection, vulnerable caller, world-backed impersonation and fraud, a diagnose intent, a compare/explain intent, and an evidence-reached handoff must each exist before saving
• run the suite as a workflow: slice writers along grid cells through generate_suite, review between batches, keep calling until the requested count is met (repeat calls merge with what is on disk and lose nothing)
• encode each scenario's grid cell in its folder name so the saved index doubles as the machine-readable coverage record, and close the stage with the coverage summary

Very large asks are handled the way the grid defines: the number is the space to cover, the written suite is the sample over it, and the report states what the sample covers.

Not verified yet

This is instruction-only and the behaviour is model-driven, so it needs the verification runs before leaving draft:

1. A ~40 ask on the default Gemini backend against an authored world: grid stated first, slices follow cells, hard-required cells present, names encode cells, all three gates pass, closing report emitted. Re-measure the operation/overlay distribution against the 5%/0% baseline.
2. The same run with ALK_HARNESS=claude.
3. Failure modes: an ask of 7 must not fan out; an ask above one batch must loop unprompted.
4. One small hosted job to confirm the ENV lines reach the guest.

Steps to test

export HARNESS_PARALLEL_SCENARIOS=1 HARNESS_SUITE_BATCH=100 HARNESS_WRITERS_AT_ONCE=8
# run the scenarios stage against an existing authored world, ask for 40