Brings the isolated pilot's uncommitted work onto a branch: per-run
resource requests (submit_v2 gateway protocol), ElasticDFT failure
diagnostics and backfill, the run-values fixes, the computational FAQ
and help pages, and the pilot provisioning and validation scripts.

Review transcripts, deployment dumps and evidence folders stay local.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QS1dbWTNbwqxXZUMEARjhm

Resolves the pilot against dev's database roles, migration runner,
proposal job queue and template editing:

- Computation store uses the api_worker role (not the superuser) and
  keeps its project-access checks; migrations create the computation
  tables unconditionally and add api_worker policies for them.
- The proposal job runner resolves objectives and domain from the
  claimed run, as the pilot's initial generation did.
- The run form keeps the pilot's typed run values and computational
  redirects, plus dev's edit-in-place and idempotency key.
- dev's "at least one planned value" check stays removed: an
  all-variable optimization run takes no point values, and the
  definition decides which inputs are required.
- Restores three pilot tests that test_*.py in .gitignore had hidden.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QS1dbWTNbwqxXZUMEARjhm

Computational tools shared the catalog tables with lab instruments and
were recognized only by conventions. Make the split explicit:

- equipment_metadata.asset_kind (physical | computational), backfilled
  for the registered RCAC installations.
- experiment_types.execution_kind, set by a trigger from the
  template's equipment, so forks of a computational template stay
  computational whichever path creates them.
- all_runs view: every run with run_kind, under each table's RLS.
- The generic run endpoint refuses computational equipment; the
  classifier prefers execution_kind over the id prefix.
- Equipment list shows Physical equipment and Computational tools
  separately; the template library badges each template's kind.
- The example RCAC worker manifest uses the api_worker role.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QS1dbWTNbwqxXZUMEARjhm

- Cache hits reuse the run that executed, never another reuse, so
  artifact lineage stays one hop after repeated identical submissions.
- An explicitly physical template is never treated as computational,
  user-created template ids never take the reserved computation_
  prefix, and the run endpoint rejects by equipment kind instead of
  guessing from the id.
- JSON object and array choices are compared as parsed JSON, not as
  Python's str() of a dict or list.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QS1dbWTNbwqxXZUMEARjhm

- The browser honours an explicitly physical template: the shared
  classifier returns non-computational for execution_kind "physical",
  and the run wizard only pre-blocks a catalog id naming a known tool
  until the template itself has loaded.
- JSON choices compare by value, so 50 matches 50.0 while true still
  does not match 1.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QS1dbWTNbwqxXZUMEARjhm

Converting to float made distinct large integers (above 2**53) equal
and could overflow; Python's own int/float comparison is exact.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QS1dbWTNbwqxXZUMEARjhm