fix: --estimate-only on a sweep estimates instead of launching (#524)

[scttfrdmn]

Closes #524.

The bug

--estimate-only was checked only inside launchSweepDetached, so a sweep that
took the foreground path launched every row while the user was asking for a
preview. That inverts the one command a user runs specifically to avoid spending
money.

Two ordinary invocations reach the foreground path:

invocation	path taken	pre-fix --estimate-only
--param-file f.yaml (default)	detached	honoured
--param-file f.yaml --no-detach --ttl 1h	foreground	launches every row
--param-file f.yaml --detach (no --max-concurrent)	foreground	launches every row

--no-detach is the documented advice for a heterogeneous sweep, because only
the foreground path detects an AMI per config (#372) — so the sweep shape most
in need of a cost preview was the one that could not get one. The bare
--detach case is subtler: maxConcurrent only gets its default inside the
auto-enable branch, so an explicit --detach leaves it at 0 and fails the
detach && maxConcurrent > 0 dispatch condition.

The fix

--estimate-only is now handled once in launchParameterSweep, above the
dispatch that chooses between the two paths, so "launches nothing" no longer
depends on which path the sweep would have taken. It is deliberately placed
before aws.NewClient: EstimateSweepCost reads only the param file, so a cost
preview needs no credentials.

The budget comparison is factored into reportSweepBudget and shared with the
real detached launch so the two cannot drift. The now-unreachable estimateOnly
block inside launchSweepDetached is removed rather than left as dead code that
would re-create the per-path shape behind the bug.

The regression test

test/e2e/tier0_sweep_estimate_test.go (Tier 0, Substrate emulator — free, no
AWS account). It asserts on the observable, not the flag: it queries EC2 for
every instance in the sweep's region and requires the set to be empty, for all
three invocations above, plus exit 0.

It deliberately does not use spawn list --state all. That flag passes
all through as a literal instance-state-name filter, which matches nothing,
so it returns [] whether or not instances exist (#527) — using it here would
have made the test unfailable, which is the same defect class the test exists to
catch.

Fails before, passes after: 2 of the 3 subtests launched 2 instances each
pre-fix. Post-fix 3/3 green, full unit suite green, full Tier 0 suite green
(158s).

Related

• Sweep dispatch silently drops --cost-limit and --ttl, then logs "Using safeguards: ttl=..." #525 — the sweep dispatch drops --cost-limit and --ttl, then logs
  Using safeguards: ttl=.... Same code path, separate defect; the pre-fix run
  of this test also reproduces it (instances came up with ttl: empty despite
  --ttl 1h).
• Unknown param-file keys silently become PARAM_* env vars — a mistyped or invented spend control caps nothing #526 — unknown param-file keys silently become PARAM_*.
• spawn list --state all silently returns nothing — "all" is passed through as a literal instance-state-name filter #527 — spawn list --state all returns [] unconditionally.

[codecov]

Codecov Report

❌ Patch coverage is 0% with 24 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
cmd/launch_sweep.go	0.00%	24 Missing ⚠️

📢 Thoughts on this report? Let us know!

[scttfrdmn]

#529 is stacked on this branch (fix for #525 — the sweep dropping --ttl/--idle-timeout/--cost-limit). Merge this one first; #529's base will retarget to main automatically.

Worth noting for review: #529 moves the CLI spend controls into paramFormat.Defaults before the --estimate-only early return this PR adds, which also fixes the estimate itself — EstimateSweepCost derives per-row hours from row ttl → defaults["ttl"] → a "1h" fallback, so before that injection --ttl 4h was priced at 1h.