e2e: scheduled flake gap-filler (default-off)

[nikw9944]

Summary of Changes

Adds an hourly schedule: trigger to the e2e workflow, guarded by a new
GitHub-hosted schedule-gate job. The gate returns proceed=true only when the
gap-filler is enabled (repo variable E2E_GAP_FILLER_ENABLED='true'), main has had
no e2e run in 6h (stale), and no e2e run is in flight on any branch (idle).
When it proceeds, the normal setup + 4-shard matrix runs on main HEAD; the
resulting run is ingested by the flake-detect system with no changes on its side.

Rationale: the e2e flake-detection corpus is starved for samples (main fails ~20%
of runs, but main runs are infrequent). This tops up samples during genuine lulls
only — the idle condition makes it yield entirely during heavy merge activity, so
it never competes for the constrained concurrent-e2e capacity.

Merges inert: the repo variable is unset by default, so scheduled ticks only run
the lightweight gate and never trigger the matrix. Enable later by setting the
variable. Design + rollout: malbeclabs/infra
docs/plans/2026-07-02-e2e-flake-gap-filler-*.md (tracker malbeclabs/infra#1615).

Only setup needed changes (a needs/if guard); the e2e matrix job already
gates on needs.setup.outputs.run-e2e, so a skipped setup skips the matrix.

Testing Verification

• actionlint .github/workflows/e2e.yml passes clean.
• node --check passes on the gate script's JS constructs.
• Non-schedule paths unchanged: setup.if evaluates true whenever
  github.event_name != 'schedule' (push/PR/dispatch run exactly as before).
• Confirmed the e2e job gates on needs.setup.outputs.run-e2e without always(),
  so proceed=false (skipped setup) cleanly skips the matrix.
• Runtime gate behavior is validated post-merge with the toggle off (GitHub runs
  schedule only on the default branch); scheduled ticks will run only
  schedule-gate and log proceed=false until enabled.