fix(forecast): project each single-stream fit from its own trajectory

[seabbs-bot]

What

Every single-stream (individual) fit was forecast by inverting its cumulative
total under exponential growth, while the joint was forecast from its own
fitted trajectory.

forecast_stream takes a stream's cut-off daily rate as the last increment of
its cumulative trajectory, and falls back to _approx_daily — inverting the
cumulative total under exponential growth over the outbreak age — when the
chain carries none. #613 added the four cumulative_* := aliases to
bvd_joint for exactly this reason, noting that the inversion "collapses
towards zero once the fitted growth rate is at or below zero, which is not what
a stream still reporting daily is doing". The single-stream composers never got
the same aliases, so every individual fit kept taking the fallback.

That is the individual-versus-joint gap visible in data/forecast_overlay.csv,
where the individual fits sit below both the joint and the observed counts at
release after release (confirmed deaths at 2026-08-24: individual median 136,
joint 315, observed 292).

What changed

• cases_only_model, deaths_only_model, confirmed_only_model and
  confirmed_deaths_only_model each expose their own stream's cumulative
  trajectory under the same un-prefixed name bvd_joint uses, so one
  _STREAM_SPEC entry keeps serving both fit kinds.
• The confirmed-case testing-onset baseline re-add moves into
  _cumulative_confirmed, shared by the joint and the confirmed-only composer,
  so cumulative_confirmed is the same quantity on both.
• _STREAM_SPEC's trajectory comment no longer says only the joint carries
  one.

No change to the joint's numbers: its cumulative_confirmed is the same
arithmetic, now through the shared helper.

Scope note

This touches src/models/joint.jl, which is hashed into every fit cache key,
so it forces a cold refit of the whole matrix.

Testing

• New :slow test item samples each of the four composers from the prior and
  asserts the trajectory key is present, is length n, never turns back on
  itself, and resolves through the same _daily_at_cutoff_any the forecaster
  uses. Red before the fix (KeyError: key :cumulative_reports not found).
• test_forecast.jl, test_boxed_captures.jl, test_cases_model.jl,
  test_confirmed_cases_model.jl, test_confirmed_deaths.jl,
  test_confirmed_cases_daily.jl, test_reported_cases_daily.jl,
  test_suspected_daily.jl, test_suspected_daily_deaths.jl: 344 passing.
  Rebased onto the feat(patch): meta-population model over the three affected provinces #412 patch-model merge and re-run there: the whole of
  test_forecast.jl, 213 passing, including feat(patch): meta-population model over the three affected provinces #412's own province-forecast
  items.

This was opened by a bot. Please ping @seabbs for any questions.

🤖 Generated with Claude Code

https://claude.ai/code/session_01XABDXNzg5H61QAxBAT2Tfx

[seabbs-review-bot]

This PR gives each of the four single-stream composers (deaths_only_model, cases_only_model, confirmed_only_model, confirmed_deaths_only_model) the same un-prefixed cumulative-trajectory := alias bvd_joint already exposes, so forecast_stream reads the cut-off daily rate directly instead of falling back to an exponential-growth inversion that collapses toward zero. The confirmed-cases logic is correctly deduplicated into a shared _cumulative_confirmed function used by both bvd_joint and confirmed_only_model, field names match the underlying submodels exactly, and the added test (paired with the prior red commit) demonstrates the behaviour change. Verdict: sound change; one documentation-placement nit.

_{Automated first pass by seabbs-review-bot (Claude sonnet), triggered by: first pass. Not a human review. Comment @seabbs-review-bot to ask for another pass: @seabbs any time, the author's agent once it has pushed changes. Add the no-review label to opt this PR out. Ping @seabbs with any questions.}

[seabbs-review-bot]

suggestion This 9-line rationale comment floats between _cumulative_confirmed and the unrelated exports_only_model docstring, so a reader meets it right before a composer that never gets a trajectory alias (:exports has an empty trajectory list in _STREAM_SPEC). It also restates, at greater length, what src/forecast.jl's _STREAM_SPEC comment already says under the trajectory: bullet, and what each of the four new := lines already says tersely at its own call site. Either delete this block (the forecast.jl doc and the per-line comments already cover it) or fold it into the _STREAM_SPEC docstring where the reader is actually resolving spec.trajectory.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 96.74%. Comparing base (54cc27c) to head (045ebfa).
⚠️ Report is 2 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #706   +/-   ##
=======================================
  Coverage   96.73%   96.74%           
=======================================
  Files          16       16           
  Lines        4536     4540    +4     
=======================================
+ Hits         4388     4392    +4     
  Misses        148      148           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.