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

docs/camdl-language-spec.md

@@ -3310,7 +3310,7 @@ sampled on a schedule. With no `output {}` block the default schedule applies;
 declare one to set the cadence or give explicit output times.
 
 > **Default schedule.** Snapshots every `1` in the model's `time_unit`, covering
-> `[min(0, t_start), t_end]` — where the window is taken from the `simulate {}`
+> `[t_start, t_end]` — where the window is taken from the `simulate {}`
 > block (or `(0, 100)` if `simulate {}` is omitted). The simulate command writes
 > the trajectory to `--output` (or stdout) and writes observation files only when
 > `--obs` / `--obs-dir` / `--obs-only` is passed.
@@ -3343,9 +3343,10 @@ metadata.json         # run provenance (see §19)
 ### 16.2 IR Mapping
 
 The trajectories block compiles to the IR `output` schedule: `every = E` →
-`OutRegular { start, step }` (start defaults to `min(0, t_start)` so the
-schedule covers anchored models with a negative `t_start`); `at = [...]` →
-`OutAtTimes`. The runtime writes the trajectory directly during simulation.
+`OutRegular { start, step }` (start defaults to `t_start` so the schedule covers
+exactly the requested window `[from, to]`, including anchored models with a
+negative `t_start`); `at = [...]` → `OutAtTimes`. The runtime writes the
+trajectory directly during simulation.
 
 Output emission is confined to `[start, simulation.t_end]`: `simulation.t_end`
 is the sole horizon authority, and output times are derived from it at emission

docs/dev/proposals/2026-06-04-schedule-unification.md

@@ -121,16 +121,27 @@ Each construct reuses it and keeps its own extension rule; the expander lowers
 - **observations** use the full core (`SchedEvery`→`ObsRegular`,
   `SchedAt`→`ObsAtTimes`). (`ObsFromData` has no frontend producer today — it is
   an IR-only stub, not a live obs `from_data` rule — so the obs surface is
-  exactly the two-arm core.) **Critical — the obs and output lowerings are NOT
-  identical:** obs `every` lowers `start = t_start` (expander.ml:3922); output
-  lowers `start = min 0.0 t_start` (expander.ml:3217). They agree at
-  `t_start = 0` (the common case) and for negative `t_start` (where
-  `min(0, t_start) = t_start`), but **diverge for `t_start > 0`** — e.g.
-  `simulate { from = 10 }` or an anchored `from` after `origin`: obs.start = 10
-  while output.start = 0. Share only the grammar rule + AST type; keep the two
-  expander lowering sites distinct with their existing `start` expressions. Do
-  **not** factor a shared `lower_schedule_core` helper — that would silently
-  shift observation times (which PGAS conditions on).
+  exactly the two-arm core.)
+
+  > **Superseded (gh#143, 40dc4f68):** the obs/output lowering **divergence**
+  > described below was a bug — output lowered `start = min 0.0 t_start`, which
+  > clamped `t_start > 0` to 0 and emitted the trajectory over `[0, t_end]`
+  > instead of `[from, to]`. Output now lowers `start = t_start`, so **both
+  > surfaces agree** in all regimes, and a shared `lower_schedule_core` helper
+  > would be safe (indeed correct) for a future Option B. The historical
+  > reasoning is kept below as the decision record; do not act on the "keep the
+  > sites distinct" instruction.
+
+  **Critical — the obs and output lowerings are NOT identical:** obs `every`
+  lowers `start = t_start` (expander.ml:3922); output lowers
+  `start = min 0.0 t_start` (expander.ml:3217). They agree at `t_start = 0` (the
+  common case) and for negative `t_start` (where `min(0, t_start) = t_start`),
+  but **diverge for `t_start > 0`** — e.g. `simulate { from = 10 }` or an
+  anchored `from` after `origin`: obs.start = 10 while output.start = 0. Share
+  only the grammar rule + AST type; keep the two expander lowering sites
+  distinct with their existing `start` expressions. Do **not** factor a shared
+  `lower_schedule_core` helper — that would silently shift observation times
+  (which PGAS conditions on).
 - **output** use the full core (`→ OutRegular` / `OutAtTimes`); `format` stays
   its own field.
 - **interventions / events** reuse only the `SchedAt` arm (explicit times).

ocaml/lib/compiler/expander.ml

@@ -5283,7 +5283,7 @@ let expand_output ctx =
   (* The output window's upper bound is no longer stored on the schedule
      (gh#143): `simulation.t_end` is the sole horizon authority, and the
      runtime derives output times from `[start, t_end]` at emission. Only
-     `start` is set here — a deliberate widening to `min(0, t_start)`. *)
+     `start` is set here — defaulting to `t_start` (see below). *)
   let t_start = match ctx.simulate with
     | None    -> 0.0
     | Some sd -> resolve_float_expr ctx sd.sim_from
@@ -5292,17 +5292,21 @@ let expand_output ctx =
     | Some { out_trajectories = Some ot; _ } -> ot.otformat
     | _ -> "tsv"
   in
-  (* Default the output schedule's start to cover the full integration
-     window. With anchored models that use `from = date(...)` before
-     `origin`, t_start is negative; an output schedule starting at 0
-     leaves no snapshots in [t_start, 0), and `--obs-only` / any
-     state-at-obs-time path (snap_at) can't find a snapshot for the
-     pre-origin observations and hard-exits. Defaulting to
-     min(0.0, t_start) preserves the existing start=0 behaviour for
-     unanchored models (t_start ≥ 0) and extends it to cover negative
-     t_start without changing the step or output cadence. (start applies
-     only to the regular schedule; `at = [...]` lists explicit times.) *)
-  let start = min 0.0 t_start in
+  (* Default the output schedule's start to the requested window origin,
+     t_start, so the trajectory covers exactly [t_start, t_end] = [from, to].
+     Two directions matter:
+       - t_start < 0 (anchored `from = date(...)` before `origin`): starting
+         at 0 would leave no snapshots in [t_start, 0), so `--obs-only` / any
+         state-at-obs-time path (snap_at) can't find a snapshot for the
+         pre-origin observations and hard-exits. start = t_start covers them.
+       - t_start > 0 (anchored `from` after `origin`, or unanchored from > 0):
+         starting at 0 would emit — and record — over [0, t_end] instead of
+         [from, to], yielding rows at times the dynamics never visited (and
+         non-monotonic time next to the t_start prologue snapshot).
+     start = t_start is correct in both directions (and trivially at t_start = 0).
+     (start applies only to the regular schedule; `at = [...]` lists explicit
+     times.) *)
+  let start = t_start in
   let times = match ctx.output_decl with
     | Some { out_trajectories = Some ot; _ } ->
       (match ot.otschedule with

ocaml/test/test_compiler.ml

@@ -800,16 +800,16 @@ let test_output_step_default () =
        Alcotest.(check (float 0.01)) "default output step" 1.0 r.Ir.step
      | _ -> Alcotest.fail "expected OutRegular schedule")
 
-(* Regression: the default output schedule must cover the full
-   integration window. With anchored models that resolve `from =
-   date(...)` to a negative t_start, the default `start = 0.0` would
-   leave [t_start, 0) without snapshots and the `--obs-only` writer
-   (and any state-at-obs-time consumer) would hard-exit with
-   "no snapshot at or before t=…" for pre-origin observations.
-   The fix: default `start = min(0.0, t_start)`. *)
+(* Regression: the default output schedule must cover exactly the requested
+   window [t_start, t_end]. With anchored models that resolve `from =
+   date(...)` to a negative t_start, a `start = 0.0` would leave [t_start, 0)
+   without snapshots and the `--obs-only` writer (and any state-at-obs-time
+   consumer) would hard-exit with "no snapshot at or before t=…" for pre-origin
+   observations. The default is `start = t_start`. *)
 
 let test_output_default_start_unanchored_stays_zero () =
-  (* Unanchored (no origin), positive t_start → output.start = 0.0 (no regression). *)
+  (* Unanchored (no origin), from = 0 → t_start = 0 → output.start = 0.0
+     (the common case; start = t_start is trivially 0 here). *)
   let src = {|
     compartments { S, I, R }
     parameters { beta : rate  gamma : rate  N0 : count  I0 : count }
@@ -857,6 +857,38 @@ let test_output_default_start_anchored_negative_t_start () =
          "output.start covers negative t_start" (-34.0) r.Ir.start
      | _ -> Alcotest.fail "expected OutRegular schedule")
 
+let test_output_default_start_anchored_positive_t_start () =
+  (* Anchored with `from = date("2020-07-01")` *after* `origin =
+     date("2020-01-01")` → t_start = 182. The default output schedule must
+     start at t_start (182), not 0: otherwise the trajectory is emitted over
+     [0, t_end] instead of the requested [from, to], producing rows at times
+     the dynamics never visited (and, alongside the t_start prologue snapshot,
+     non-monotonic time). This is the anchored-positive counterpart of the
+     negative-t_start case above. *)
+  let src = {|
+    time_unit = 'days
+    origin = date("2020-01-01")
+    compartments { S, I, R }
+    parameters { beta : rate  gamma : rate  N0 : count  I0 : count }
+    let N = S + I + R
+    transitions {
+      infection : S --> I  @ beta * S * I / N
+      recovery  : I --> R  @ gamma * I
+    }
+    init { S = N0 - I0  I = I0 }
+    simulate { from = date("2020-07-01")  to = date("2020-07-12") }
+  |} in
+  match Compiler.compile ~name:"test_output_anchored_pos_start" src with
+  | Error e -> Alcotest.failf "compile failed: %s" e
+  | Ok m ->
+    Alcotest.(check (float 1e-9)) "t_start" 182.0 m.Ir.simulation.Ir.t_start;
+    (match m.Ir.output.Ir.times with
+     | Ir.OutRegular r ->
+       Alcotest.(check (float 1e-9))
+         "output.start = t_start (window is [from, to], not [0, to])"
+         182.0 r.Ir.start
+     | _ -> Alcotest.fail "expected OutRegular schedule")
+
 (* Output trajectory customization (Phase 1): the trajectories block accepts
    `every = E` (regular cadence) and `at = [...]` (explicit times), mirroring
    the observation schedule surface, plus `format = NAME`. *)
@@ -919,13 +951,14 @@ let test_output_every_and_at_conflict () =
   | Error _ -> ()  (* expected: conflicting schedule rejected *)
   | Ok _ -> Alcotest.fail "expected error: every and at are mutually exclusive"
 
-(* A.2 guard: observations and output share the `schedule_core` grammar/AST
-   but their expander lowerings are NOT identical — obs `every` lowers
-   start = t_start, output `every` lowers start = min(0, t_start). For
-   t_start > 0 they diverge (obs.start = t_start, output.start = 0). Pin it
-   so a future shared schedule_core lowering helper can't silently shift
-   observation times (which PGAS conditions on). *)
-let test_obs_output_start_divergence () =
+(* A.2 guard: observations and output share the `schedule_core` grammar/AST,
+   and their expander lowerings now agree — both lower `every` to
+   start = t_start. For t_start > 0 (here from = 10 'days) obs.start and
+   output.start both equal 10, so the trajectory covers [from, to] and the
+   observation times (which PGAS conditions on) line up with it. Pin the
+   agreement so a future shared schedule_core lowering helper stays honest and
+   the min(0, t_start) window bug can't silently return. *)
+let test_obs_output_start_agrees () =
   let src = {|
     compartments { S, I, R }
     parameters { beta : rate  gamma : rate  rho : rate  N0 : count  I0 : count }
@@ -951,14 +984,14 @@ let test_obs_output_start_divergence () =
   | Ok m ->
     (match m.Ir.output.Ir.times with
      | Ir.OutRegular r ->
-       Alcotest.(check (float 1e-9)) "output.start = min(0,t_start) = 0" 0.0 r.Ir.start
+       Alcotest.(check (float 1e-9)) "output.start = t_start = 10" 10.0 r.Ir.start
      | _ -> Alcotest.fail "expected OutRegular output schedule");
     (match m.Ir.observations with
      | om :: _ ->
        (match om.Ir.emit_schedule with
         | Some (Ir.ObsRegular r) ->
           Alcotest.(check (float 1e-9))
-            "obs.start = t_start = 10 (NOT min(0,t_start))" 10.0 r.Ir.start
+            "obs.start = t_start = 10 (agrees with output.start)" 10.0 r.Ir.start
         | _ -> Alcotest.fail "expected ObsRegular emit_schedule")
      | [] -> Alcotest.fail "expected an observation model")
 
@@ -9421,13 +9454,15 @@ let () =
         `Quick test_output_default_start_unanchored_stays_zero;
       Alcotest.test_case "anchored t_start<0 → output.start covers full integration window"
         `Quick test_output_default_start_anchored_negative_t_start;
+      Alcotest.test_case "anchored t_start>0 → output.start = t_start (window is [from, to])"
+        `Quick test_output_default_start_anchored_positive_t_start;
       Alcotest.test_case "every = E → OutRegular step" `Quick test_output_every_explicit;
       Alcotest.test_case "every = 0.5 → sub-unit cadence" `Quick test_output_every_subunit;
       Alcotest.test_case "at = [...] → OutAtTimes" `Quick test_output_at_times;
       Alcotest.test_case "format = parquet" `Quick test_output_format_parquet;
       Alcotest.test_case "every and at conflict → error" `Quick test_output_every_and_at_conflict;
-      Alcotest.test_case "obs.start=t_start vs output.start=0 (A.2 lowering divergence guard)"
-        `Quick test_obs_output_start_divergence;
+      Alcotest.test_case "obs.start = output.start = t_start (A.2 lowering agreement guard)"
+        `Quick test_obs_output_start_agrees;
       Alcotest.test_case "stratified obs header emits (dim,level) stratum + serde round-trip"
         `Quick test_stratified_observation_emits_stratum;
     ];