fix(expander): default output window starts at t_start, not min(0, t_start) (gh#143)

[vsbuffalo]

What

The default trajectory output window started at min(0, t_start) instead
of t_start. For any model with t_start > 0 — an anchored from later
than origin, or an unanchored from > 0 — and **no explicit output
block, this forced the window start to 0, so the trajectory was emitted
(and recorded) over [0, t_end] instead of the requested [from, to].

Why it was wrong

min(0, t_start) was the gh#143 fix for the negative-t_start case (a
from before origin): starting at 0 leaves no snapshots in [t_start, 0),
so --obs-only / snap_at can't find a pre-origin snapshot and hard-exits.
But min(0, t_start) equals t_start for t_start < 0 and clamps to 0 for
t_start > 0 — it fixed the negative direction and silently broke the
positive one. start = t_start is correct in both directions (and
trivially at t_start = 0), subsuming the gh#143 negative-t_start fix
rather than replacing it.

Reproduction

Minimal model (origin = 2020-01-01, from = 2020-07-01,
to = 2020-07-12 → t_start = 182, t_end = 193), no output block:

• before: camdl simulate m.camdl --dates --stdout emits 194 data rows
  with non-monotonic time (182, then 1, 2, …, 193).
• after: 12 data rows over [182, 193], strictly increasing time.

Tests

TDD, red → green. New output_schedule unit test for the anchored
t_start > 0 case:

# RED (current code)
[FAIL] anchored t_start>0 → output.start = t_start
  Expected: 182
  Received: 0

# GREEN (after fix)
Test Successful. 12 tests run.

The former A.2 obs/output divergence guard is repurposed as an
agreement guard: both surfaces now lower start = t_start, so they line
up (and a future shared schedule_core lowering helper would be safe).

Golden impact

None. No committed golden has t_start > 0 (all are unanchored or use
at = [...]), verified by scanning all 108 committed *.ir.json. No IR
golden or trajectory baseline changes.

Spec §16 synced to the fixed reality ([t_start, t_end]; start defaults to
t_start).

Full make test green.

Follow-up (not in this PR)

The landed proposal docs/dev/proposals/2026-06-04-schedule-unification.md
(§A.2) still records the old divergence and instructs a future Option-B
implementer not to factor a shared lowering helper "because it would shift
observation times." That premise is now inverted — obs and output agree — so
that guidance is stale. Left untouched here to keep this PR to one logical
change; the live regression guard (the renamed agreement test) already pins
the new invariant.