docs: migrate tutorial notebooks off fit-time aggregate= (sweep PR-B)

[igerber]

Summary

• Completes the fit-time aggregate= teaching sweep (notebook half; follows docs: migrate narrative RST off fit-time aggregate= (sweep PR-A) #763): 16 analytical fit/configuration sites across 02_staggered_did, 16_survey_did (including its 200-draw Monte-Carlo loop), 26_composition_drift_calibration (native seam + the two balance-adapter sites, where the kwarg was not load-bearing), 16_wooldridge_etwfe, 17_brand_awareness_survey, and 24_staggered_vs_collapsed_power now fit plain and aggregate post-fit via results.aggregate(type=...).
• Legacy event_study_effects/group_effects dict reads are rebound to the containers with keyed lookups that preserve each notebook's missing-event-time fallbacks (the compared grids differ in size, so positional pairing would misalign rows); cells that printed the fit-time summary() aggregation blocks print the container summaries instead, so no rendered teaching content is lost; plot_event_study and compute_honest_did consume the containers directly.
• Deliberate keeps, each annotated in-notebook: the BOOTSTRAPPED fits (02 cell 20, both 09_real_world_examples case studies) retain fit-time aggregate= with comments naming it the documented until-4.0 exception (post-fit recompute levels raise on a bootstrapped fit; 'simple' relays), with the expected FutureWarning visible in the refreshed outputs and acknowledged in adjacent markdown; 08_triple_diff's staggered DDD keeps fit-time as the CANONICAL route (rows M-140/M-141 - no post-fit successor exists on its container).
• tests/test_t26_composition_drift_calibration_drift.py mirrors the tutorial-26 migration at all three of its sites (the fit_survey_cs mirror, the legacy-field read, and the adapter-parity call).
• TODO.md: the sweep row is retired; the kept teaching surfaces stay deviation-grep-resolvable by moving onto the staggered-DDD post-fit-aggregate() row and the diagnostic_report-admission row Locations.
• All 8 edited notebooks were re-executed with outputs written back, under DIFF_DIFF_BACKEND=python (the backend both notebook CI jobs and RTD-rendered outputs use) on this worktree's kernel. Diff-size note: the ~6,700 insertions are dominated by refreshed outputs plus nbformat's canonical normalization of two notebooks previously stored as compact single-string-source JSON; the reviewable source delta is the cell-source changes described above.

Methodology references (required if estimator / math changes)

• Method name(s): N/A - documentation/tutorial-only; no estimator or math changes. Post-fit aggregation reuses the same aggregation machinery the fit-time path called (Callaway & Sant'Anna 2021 group-time -> aggregated parameters, mirroring R did's post-estimation aggte()).
• Paper / source link(s): N/A
• Any intentional deviations from the source (and why): None

Validation

• Tests added/updated: tests/test_t26_composition_drift_calibration_drift.py (three mirror sites migrated; one assertion now reads the post-fit container). Locally: pytest --nbmake over all 8 edited notebooks passes (8/8, DIFF_DIFF_BACKEND=python); tests/test_t24_staggered_vs_collapsed_power_drift.py (incl. slow-marked) and tests/test_t26_* pass; a scoped sweep (grep -rn "aggregate=" docs/*.rst docs/api/*.rst docs/tutorials/*.ipynb) confirms the only remaining fit-time sites are the sanctioned inventory (bootstrap exceptions, canonical DDD, blocked report API pages, already-correct prose).
• Backtest / simulation / notebook evidence (if applicable): refreshed committed outputs for all 8 notebooks; migrated cells render warning-free, kept-exception cells render the expected FutureWarning.

Security / privacy

• Confirm no secrets/PII in this PR: Yes

[github-actions]

Overall assessment: ✅ Looks good

Executive summary

• No unmitigated P0/P1 findings.
• No estimator, weighting, variance, or identification logic changed.
• Post-fit aggregation preserves event-time alignment and survey inference.
• Bootstrap and staggered-DDD exceptions are documented and tracked.
• Independent tests could not run because this environment lacks pytest/NumPy.

Methodology

• Severity: P3 informational
  Impact: None. Changes use the existing Callaway–Sant’Anna aggregation machinery and preserve HonestDiD covariance/provenance (docs/tutorials/17_brand_awareness_survey.ipynb:L1058).
  Concrete fix: None required.

Code Quality

• Severity: None
  Impact: Keyed lookups correctly avoid mismatching differing event-time grids.
  Concrete fix: None.

Performance

• Severity: None
  Impact: The Monte Carlo loop still performs one event-study aggregation per fit; no material regression identified.
  Concrete fix: None.

Maintainability

• Severity: None
  Impact: Tutorials now consistently use public post-fit containers.
  Concrete fix: None.

Tech Debt

• Severity: P3 informational
  Impact: Bootstrap re-aggregation and staggered-DDD post-fit aggregation remain known limitations, tracked in TODO.md:L24 and TODO.md:L54.
  Concrete fix: None required in this PR.

Security

• Severity: P3
  Impact: Refreshed outputs contain machine-specific absolute paths and a local username, but no credentials or secrets (docs/tutorials/17_brand_awareness_survey.ipynb:L433).
  Concrete fix: Optionally sanitize warning filenames or execute notebooks from a neutral workspace path.

Documentation/Tests

• Severity: None
  Impact: All changed notebooks contain execution counts and no stored error outputs; the t26 mirror test uses the new container (tests/test_t26_composition_drift_calibration_drift.py:L320).
  Concrete fix: None.

[github-actions]

🔁 AI review rerun (requested by @igerber)

Head SHA: 75c8b459b7e6dbf11016d4f3afefcb158deff638

---

Overall assessment: ✅ Looks good

Executive summary

• No unmitigated P0/P1 findings.
• No estimator, weighting, variance, or identification logic changed.
• Post-fit containers preserve keyed event-time alignment and HonestDiD covariance.
• Bootstrap and staggered-DDD exceptions match documented contracts and TODO.md.
• The prior machine-path P3 is resolved; no secrets found.
• Tests could not run because NumPy/pytest are unavailable; all notebooks have execution counts and no stored errors.

Methodology

• Severity: P3 informational
  Impact: Migrations use the existing Callaway–Sant’Anna post-fit aggregation contract, including survey inference and HonestDiD consumption (docs/tutorials/17_brand_awareness_survey.ipynb:L1058, L1198).
  Concrete fix: None.

Code Quality

• Severity: None
  Impact: Keyed lookups correctly handle unequal event-time grids (docs/tutorials/02_staggered_did.ipynb:L1494, L2102).
  Concrete fix: None.

Performance

• Severity: None
  Impact: No material computation regression identified.
  Concrete fix: None.

Maintainability

• Severity: None
  Impact: Analytical tutorials now consistently use public post-fit containers.
  Concrete fix: None.

Tech Debt

• Severity: P3 informational
  Impact: Bootstrap re-aggregation and staggered-DDD aggregation remain explicitly tracked (TODO.md:L24, L54).
  Concrete fix: None required here.

Security

• Severity: P3 informational
  Impact: The previous machine-specific path disclosure is resolved; refreshed warnings use <ipython-cell>.py.
  Concrete fix: None.

Documentation/Tests

• Severity: None
  Impact: The t26 mirror test uses post-fit aggregation (tests/test_t26_composition_drift_calibration_drift.py:L320); notebook metadata contains no execution errors.
  Concrete fix: None.