Review fixes:
- Add edge case validation in _compute_flci (se > 0, 0 < alpha < 1)
- Improve significance_stars docstring explaining partial identification
- Standardize error messages to include parameter values (M, Mbar, alpha)
- Make LP solver method configurable in _solve_bounds_lp
- Add clarifying comment about constraint matrix design for pre+post periods
- Improve CallawaySantAnna error message with actionable guidance

Notes:
- #4 (sensitivity_plot export) was verified as valid - function exists at
  honest_did.py:1437
- #1 (pre-period effects) verified correct - LP optimization covers all
  periods but only post-periods contribute to objective function

Revised review reflects:
- #1, #4 verified as non-issues (correct by design)
- #3, #5, #6, #8, #13 addressed in commit e40d6b4
- Updated recommendation to approve and merge
- Remaining items are low-priority style suggestions for future PRs

Phase 2 silent-failures audit — axis-G (backend parity). Closes the
coverage gap the audit flagged in three Rust-backed solver surfaces.
Test-only PR; any discovered divergences are marked `xfail(strict=True)`
and logged to `TODO.md` as P1 follow-ups rather than fixed in-scope.

Finding #21 — `solve_ols` skip-rank-check parity (`linalg.py:369-373,
597-639`): three parity tests in `TestSolveOLSSkipRankCheckParity`
covering mixed-scale columns (norm ratio > 1e6), near-singular full-rank
(cond > 1e10), and rank-deficient collinear designs under
`skip_rank_check=True` on HC1. Backends agree on fitted values within
`rtol=1e-6, atol=1e-8`. All pass; no Rust-side code change needed.

Finding #22 — `compute_synthetic_weights` parity (`utils.py:1134-1199`):
three parity tests in `TestSyntheticWeightsBackendParity`. Near-singular
`Y'Y` passes at `atol=1e-7`; extreme Y scale (1e9) and lambda_reg
variations are `xfail(strict=True)` with a baselined ~15-80% weight
divergence. Root cause: Rust path is Frank-Wolfe, Python fallback is
projected gradient descent (`utils.py:1228`) — same QP, different
simplex vertices under near-degenerate inputs.

Finding #23 — TROP Rust grid-search + bootstrap parity
(`trop_global.py:688-750, 966-1006`): two parity tests in
`TestTROPRustEdgeCaseParity`, `@pytest.mark.slow` class-level. Both
`xfail(strict=True)`: grid-search ATT on rank-deficient Y (~6%
divergence), bootstrap SE under `seed=42` (~28% divergence, RNG
backend mismatch — Rust `rand` crate vs numpy `default_rng`).

Plan governance:
- Per `feedback_ci_reviewer_pattern_checks`, greped adjacent Rust
  entry points (`_solve_ols_rust`, `_rust_synthetic_weights`,
  `_rust_loocv_grid_search_global`, `_rust_bootstrap_trop_variance_global`);
  no additional silent-fallback surfaces identified.
- Per plan Non-goal #4, did not open an axis-H finding on TROP's
  `seed=None → 0` substitution at `trop_global.py:994` (out of scope).
- No behavioral changes, no warnings, no REGISTRY changes, no flags.

TODO.md logs three P1 follow-up entries: algorithmic unification for
`compute_synthetic_weights` (FW vs PGD), TROP grid-search divergence on
rank-deficient Y, TROP bootstrap RNG unification.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Closes BR/DR foundation gap #4 (real-dataset validation) from the
external-positioning gap list in ``project_br_dr_foundation.md``.

Validation artifact:

- ``docs/validation/validate_br_dr_canonical.py`` runs BusinessReport
  / DiagnosticReport on Card-Krueger (1994), mpdta (Callaway-Sant'Anna
  2021 benchmark), and Castle Doctrine (Cheng-Hoekstra 2013 under
  both CS and SA), dumping summary + full_report + selected to_dict
  blocks for each.
- ``docs/validation/br_dr_canonical_validation.md`` is the regenerable
  raw output.
- ``docs/validation/br_dr_canonical_findings.md`` is the hand-written
  synthesis: direction / verdict / sensitivity tier all match canonical
  interpretations, with two small wording bugs surfaced and fixed in
  this PR and two larger gaps queued as follow-up (SA HonestDiD
  applicability, target-parameter disambiguation).

Wording fixes:

1. Treatment-label capitalization. ``str.capitalize()`` lowercased
   every character after the first, flattening embedded abbreviations
   (``"the NJ minimum-wage increase"`` → ``"The nj minimum-wage
   increase"``) and proper-noun phrases (``"Castle Doctrine law
   adoption"`` → ``"Castle doctrine law adoption"``). Replaced with a
   ``_sentence_first_upper`` helper that preserves user-supplied
   casing.

2. ``breakdown_M == 0`` phrasing. The HonestDiD fragile sentence
   quoted ``{breakdown_M:.2g}x the pre-period variation``, which
   renders as a degenerate ``0x`` on the exact-zero case surfaced by
   Cheng-Hoekstra. At ``breakdown_M <= 0.05`` (covers 0 and near-zero
   values), both BR's summary and DR's overall_interpretation now say
   "includes zero even at the smallest parallel-trends violations on
   the sensitivity grid" instead.

Tests: 5 new regressions in
``TestCanonicalValidationSurfaceFixes`` covering both fixes + three
boundary cases (exact zero, small positive, normal fragile value).

Not in scope: Favara-Imbs (dCDH reversible-treatment dataset not
bundled), ImputationDiD / TwoStageDiD on canonical data (needed to
exercise the R42 untreated-outcome FE assumption branch on real
data), SA HonestDiD applicability gap. All tracked in the findings
doc for follow-up.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Close BR/DR gap #4: canonical-dataset regression guards + wording fixes

…nuousDiD prerequisite list as profile-side screening + add first_treat caveat

P1 (the five profile-derived facts are not the "full" gate set):
Reviewer correctly noted that calling
`{has_never_treated, treatment_varies_within_unit==False,
is_balanced, no duplicate_unit_time_rows alert, dose_min > 0}` the
"full ContinuousDiD pre-fit gate set" overreaches. `profile_panel`
only sees the four columns it accepts and CANNOT see the separate
`first_treat` column that `ContinuousDiD.fit()` consumes. Verified
against `continuous_did.py:230-360`: `fit()` additionally rejects
NaN/inf/negative `first_treat`, drops units with `first_treat > 0`
AND `dose == 0`, and force-zeroes `first_treat == 0` rows whose
`dose != 0` with a `UserWarning`. A panel that passes all five
profile-side checks can still surface warnings, drop rows, or raise
at fit time depending on the `first_treat` column the caller
supplies.

Reframed the wording in five surfaces from "full gate set" to
"profile-side screening checks" with an explicit caveat that the
checks are necessary-but-not-sufficient and that `ContinuousDiD.fit()`
applies separate `first_treat` validation:

- `diff_diff/profile.py` `TreatmentDoseShape` docstring (now spells
  out the screening framing explicitly + lists the `first_treat`
  validations that fit() applies).
- `diff_diff/profile.py` `_compute_treatment_dose` helper docstring
  (aligned with public contract: most fields descriptive,
  `dose_min > 0` is one of the screening checks).
- `diff_diff/guides/llms-autonomous.txt` §2 field reference (rewrote
  the multi-paragraph block to describe screening + first_treat
  caveat).
- `diff_diff/guides/llms-autonomous.txt` §4.7 (continuous design
  feature paragraph: screening checks + necessary-not-sufficient
  language + pointer to §2).
- `diff_diff/guides/llms-autonomous.txt` §5.2 worked example
  reasoning chain (rewrote step 2 to call out screening +
  first_treat caveat; clarified counter-example #4 that
  `P(D=0) > 0` is required under BOTH `control_group="never_treated"`
  and `"not_yet_treated"`, not just default).
- `CHANGELOG.md` Unreleased entry.
- `ROADMAP.md` AI-Agent Track.

P2 (test coverage for the missing `first_treat` caveat):
Added a content-stability assertion in `tests/test_guides.py`:
`assert "first_treat" in text` so the autonomous guide cannot
silently drop the explicit `first_treat` validation caveat.

P3 (helper / test-name inconsistency with public contract):
Renamed `test_treatment_dose_does_not_gate_continuous_did` to
`test_treatment_dose_descriptive_fields_supplement_existing_gates`
and rewrote its docstring to match the now-honest public contract
("most fields descriptive distributional context that supplements
the existing top-level screening checks"). The test body still
asserts the same two things — `treatment_varies_within_unit` fires
True on `0,0,d,d` paths and `has_never_treated` is independent of
`has_zero_dose` — both of which remain accurate.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…nuousDiD prerequisite list as profile-side screening + add first_treat caveat

P1 (the five profile-derived facts are not the "full" gate set):
Reviewer correctly noted that calling
`{has_never_treated, treatment_varies_within_unit==False,
is_balanced, no duplicate_unit_time_rows alert, dose_min > 0}` the
"full ContinuousDiD pre-fit gate set" overreaches. `profile_panel`
only sees the four columns it accepts and CANNOT see the separate
`first_treat` column that `ContinuousDiD.fit()` consumes. Verified
against `continuous_did.py:230-360`: `fit()` additionally rejects
NaN/inf/negative `first_treat`, drops units with `first_treat > 0`
AND `dose == 0`, and force-zeroes `first_treat == 0` rows whose
`dose != 0` with a `UserWarning`. A panel that passes all five
profile-side checks can still surface warnings, drop rows, or raise
at fit time depending on the `first_treat` column the caller
supplies.

Reframed the wording in five surfaces from "full gate set" to
"profile-side screening checks" with an explicit caveat that the
checks are necessary-but-not-sufficient and that `ContinuousDiD.fit()`
applies separate `first_treat` validation:

- `diff_diff/profile.py` `TreatmentDoseShape` docstring (now spells
  out the screening framing explicitly + lists the `first_treat`
  validations that fit() applies).
- `diff_diff/profile.py` `_compute_treatment_dose` helper docstring
  (aligned with public contract: most fields descriptive,
  `dose_min > 0` is one of the screening checks).
- `diff_diff/guides/llms-autonomous.txt` §2 field reference (rewrote
  the multi-paragraph block to describe screening + first_treat
  caveat).
- `diff_diff/guides/llms-autonomous.txt` §4.7 (continuous design
  feature paragraph: screening checks + necessary-not-sufficient
  language + pointer to §2).
- `diff_diff/guides/llms-autonomous.txt` §5.2 worked example
  reasoning chain (rewrote step 2 to call out screening +
  first_treat caveat; clarified counter-example #4 that
  `P(D=0) > 0` is required under BOTH `control_group="never_treated"`
  and `"not_yet_treated"`, not just default).
- `CHANGELOG.md` Unreleased entry.
- `ROADMAP.md` AI-Agent Track.

P2 (test coverage for the missing `first_treat` caveat):
Added a content-stability assertion in `tests/test_guides.py`:
`assert "first_treat" in text` so the autonomous guide cannot
silently drop the explicit `first_treat` validation caveat.

P3 (helper / test-name inconsistency with public contract):
Renamed `test_treatment_dose_does_not_gate_continuous_did` to
`test_treatment_dose_descriptive_fields_supplement_existing_gates`
and rewrote its docstring to match the now-honest public contract
("most fields descriptive distributional context that supplements
the existing top-level screening checks"). The test body still
asserts the same two things — `treatment_varies_within_unit` fires
True on `0,0,d,d` paths and `has_never_treated` is independent of
`has_zero_dose` — both of which remain accurate.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…light checks as standard-workflow predictions, not estimator gates

Reviewer correctly noted that calling
{has_never_treated, treatment_varies_within_unit==False,
is_balanced, no duplicate_unit_time_rows alert, dose_min > 0}
the "screening checks" / "necessary" gates of `ContinuousDiD`
overstates the contract. `ContinuousDiD.fit()` keys off the
separate `first_treat` column (which `profile_panel` does not see),
defines never-treated controls as `first_treat == 0` rows,
force-zeroes nonzero `dose` on those rows with a `UserWarning`,
and rejects negative dose only among treated units `first_treat > 0`
(see `continuous_did.py:276-327` and `:348-360`).

Two of the five checks (`has_never_treated`, `dose_min > 0`) are
first_treat-dependent: agents who relabel positive- or negative-dose
units as `first_treat == 0` trigger the force-zero coercion path
with a `UserWarning` and may still fit panels that fail those
preflights, with the methodology shifting. The other three
(`treatment_varies_within_unit`, `is_balanced`, duplicate-row
absence) are real fit-time gates that hold regardless of how
`first_treat` is constructed.

Reframed every wording site to call these "standard-workflow
preflight checks" — predictive when the agent derives `first_treat`
from the same dose column passed to `profile_panel`, but not the
estimator's literal contract:

- `diff_diff/profile.py` `TreatmentDoseShape` docstring (rewrote
  the multi-paragraph block; explicit standard-workflow definition
  + per-check first_treat dependency map + force-zero coercion
  caveat).
- `diff_diff/profile.py` `_compute_treatment_dose` helper docstring
  (already brief; stays consistent).
- `diff_diff/guides/llms-autonomous.txt` §2 field reference (long
  rewrite covering the standard-workflow framing + override paths).
- `diff_diff/guides/llms-autonomous.txt` §4.7 opening bullet +
  trailing paragraph (both updated; opening bullet now spells out
  which of the five checks are first_treat-dependent vs. hard
  fit-time stops; trailing paragraph promotes the standard-
  workflow framing).
- `diff_diff/guides/llms-autonomous.txt` §5.2 reasoning chain step
  2 (rewrote the gate-checking paragraph; counter-example #4
  expanded to enumerate (a) supply matching first_treat and accept
  rejection, (b) deliberate relabel + coercion, (c) different
  estimator; counter-example #5 distinguishes negative-dose
  treated-unit rejection from never-treated coercion).
- `CHANGELOG.md` Wave 2 entry (matches the new framing).
- `ROADMAP.md` AI-Agent Track building block (matches).

Test coverage:
- Renamed assertion messages in
  `test_treatment_dose_descriptive_fields_supplement_existing_gates`
  and `test_treatment_dose_min_flags_negative_dose_continuous_panels`
  to remove "authoritative gate" phrasing; reframed as "standard-
  workflow preflight" assertions consistent with the corrected docs.
- Added `test_negative_dose_on_never_treated_coerces_not_rejects`
  in `tests/test_continuous_did.py::TestEdgeCases` covering the
  reviewer's specific request: never-treated rows with NEGATIVE
  nonzero dose must coerce (with `UserWarning`) rather than raise
  the treated-unit negative-dose error. Sister to the existing
  `test_nonzero_dose_on_never_treated_warns` which covers the
  positive-dose case.

Rebased onto origin/main during this round (no conflicts beyond
prior CHANGELOG resolutions; main advanced 19 commits).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…s-fallback wording; correct duplicate-row "fit-time stop" claim

P1 (relabel-to-manufacture-controls misframing):
Round 11 introduced wording across the guide, profile docstring,
CHANGELOG, ROADMAP, and test docstrings that presented intentional
`first_treat == 0` relabeling of nonzero-dose units as an
"option" / "fallback" for fitting `ContinuousDiD` when the
profile-side preflights (`has_never_treated`, `dose_min > 0`)
fail. REGISTRY does not document this as a routing option, and the
estimator still requires actual `P(D=0) > 0` because Remark 3.1
lowest-dose-as-control is not yet implemented. The force-zero
coercion at `continuous_did.py:311-327` is implementation behavior
for INCONSISTENT inputs (e.g., user accidentally passes nonzero
dose on a never-treated row), not a methodology fallback.

Reworded every site to remove the relabeling-as-option framing and
replace it with the registry-documented fixes when (1) or (5)
fails: re-encode the treatment column to a non-negative scale that
contains a true never-treated group, or route to a different
estimator (`HeterogeneousAdoptionDiD` for graded-adoption panels;
linear DiD with the treatment as a continuous covariate). Every
remaining "manufacture controls" mention in the guide, profile,
and tests is now an explicit anti-recommendation ("do not relabel
... to manufacture controls"). Updated:

- `diff_diff/profile.py` `TreatmentDoseShape` docstring (item (1):
  "not an opportunity to relabel ..."; item (5): coercion is
  "implementation behavior for inconsistent inputs, not a
  methodological fallback").
- `diff_diff/guides/llms-autonomous.txt` §2 field reference (the
  When-(1)-or-(5)-fails paragraph names re-encode + alternative
  estimator only; explicit anti-relabel warning).
- `diff_diff/guides/llms-autonomous.txt` §4.7 opening bullet +
  trailing paragraph (consolidated; opening bullet drops the
  relabel-as-fallback framing; trailing paragraph trimmed to a
  pointer to §2).
- `diff_diff/guides/llms-autonomous.txt` §5.2 step 2 + counter-
  example #4 + counter-example #5 (relabel-as-option language
  removed; explicit "do not relabel" callouts; counter-example #4
  options trimmed to (a) re-encode and (b) different estimator).
- `CHANGELOG.md` (relabel-as-option clause removed; replaced with
  re-encode / different-estimator framing).
- `ROADMAP.md` (same).
- `tests/test_profile_panel.py` two test docstrings (relabel-as-
  workflow language removed).

P2 (duplicate-row "hard fit-time stop" misclaim):
Round 11 wording said "duplicate-row failures are hard fit-time
stops" — incorrect. `_precompute_structures` at
`continuous_did.py:818-823` silently overwrites with last-row-wins,
no exception raised. Reworded as "hard preflight veto: the agent
must deduplicate before fit because `ContinuousDiD` otherwise uses
last-row-wins, no fit-time exception" in profile.py docstring,
guide §4.7 opening bullet, and §5.2 step 2 (now defers to §2 for
the breakdown). The previously-correct §2 description of the
silent-coerce path is preserved.

Length housekeeping:
The round-11 round-12 expansion pushed `llms-autonomous.txt`
above `llms-full.txt`, breaking `test_full_is_largest`. Trimmed
~2.7KB by consolidating the §4.7 trailing paragraph + §5.2 step 2
trailing block to point at §2's full breakdown rather than
duplicating the per-check semantics. autonomous: 65364 chars,
full: 66058 chars.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

… first_treat from dose" framing; add PanelProfile backward-compat defaults; fix test_continuous_did docstring

P1 (canonical ContinuousDiD setup vs. derive-from-dose framing):
Round 12 introduced a "standard workflow" description across the
guide, profile docstring, CHANGELOG, ROADMAP, and test docstrings
that said agents derive `first_treat` from the same dose column
passed to `profile_panel`. Reviewer correctly noted this conflicts
with the actual ContinuousDiD contract (`continuous_did.py:222-228`,
`prep_dgp.py:970-993`, `docs/methodology/continuous-did.md:65-73`):
the canonical setup uses a **time-invariant per-unit dose** `D_i`
and a **separate `first_treat` column** the caller supplies — the
dose column has no within-unit time variation in this setup, so it
cannot encode timing. An agent following the rejected framing would
either build a `0,0,d,d` path (which `fit()` rejects) or keep a
valid constant-dose panel (in which case the dose column carries no
timing information).

Reworded every site to drop the derive-from-dose framing and
replace with the canonical setup. The five facts on the dose column
remain predictive of `fit()` outcomes BECAUSE the canonical
convention ties `first_treat == 0` to `D_i == 0` and treated units
carry their constant dose across all periods — so `has_never_treated`
proxies `P(D=0) > 0` and `dose_min > 0` predicts the strictly-
positive-treated-dose requirement, without any "derivation" of
`first_treat` from the dose column. Updated:

- `diff_diff/profile.py` `TreatmentDoseShape` docstring (rewrote
  the multi-paragraph block to use the canonical-setup framing
  and added an explicit "agent must validate `first_treat`
  independently" note).
- `diff_diff/guides/llms-autonomous.txt` §2 field reference.
- `diff_diff/guides/llms-autonomous.txt` §4.7 opening bullet.
- `diff_diff/guides/llms-autonomous.txt` §5.2 reasoning chain
  step 2 + counter-examples #4 and #5 (now describe the
  canonical setup rather than a derive-from-dose workflow).
- `CHANGELOG.md` Wave 2 entry.
- `ROADMAP.md` AI-Agent Track building block.
- `tests/test_profile_panel.py` `test_treatment_dose_min_flags
  _negative_dose_continuous_panels` docstring/comments.

P2 (PanelProfile direct-construction backward compat):
Wave 2 added `outcome_shape` and `treatment_dose` to PanelProfile
without defaults, breaking direct `PanelProfile(...)` calls that
predate Wave 2. Made both fields default to `None` (moved them to
the end of the field list; both are `Optional[...]`). Added
`test_panel_profile_direct_construction_without_wave2_fields`
asserting that direct construction without the new fields succeeds
and yields `None` defaults that serialize correctly through
`to_dict()`.

P3 (test_continuous_did.py docstring overstating sanction):
The new `test_negative_dose_on_never_treated_coerces_not_rejects`
docstring said the contract "lets agents legally relabel
negative-dose units as `first_treat == 0` to coerce them away."
Reworded as observed implementation behavior for inconsistent
inputs, NOT a sanctioned routing option — the test locks in the
coercion contract while the autonomous guide §5.2 explicitly tells
agents not to use this path methodologically.

Length invariant maintained: autonomous (65748 chars) < full
(66031 chars); `test_full_is_largest` still passes (compares
character count, not byte count, so on-disk size with UTF-8
multi-byte characters differs from the assertion target).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…fixes" overclaim for ContinuousDiD recoding

P1 (overclaiming registry endorsement of recoding):
Reviewer correctly noted the round-13/14 wording across the
public-facing surfaces called re-encoding the treatment column a
"registry-documented fix" / "documented option" / "documented
fallback". REGISTRY only documents the `P(D=0) > 0` requirement
and explicitly notes Remark 3.1's lowest-dose-as-control fallback
is NOT implemented in this library. Re-encoding is an agent-side
preprocessing choice that the registry neither endorses nor
forbids — calling it "registry-documented" was an over-claim.

Reworded twelve sites to drop the "documented" framing:
- `diff_diff/profile.py` `TreatmentDoseShape` docstring (items
  (1) and (5)).
- `diff_diff/guides/llms-autonomous.txt` §2 field reference
  When-(1)-or-(5)-fails paragraph.
- `diff_diff/guides/llms-autonomous.txt` §4.7 opening bullet
  trailing language.
- `diff_diff/guides/llms-autonomous.txt` §4.7 trailing paragraph
  (consolidated to a pointer at §2; reduced redundancy).
- `diff_diff/guides/llms-autonomous.txt` §5.2 reasoning chain
  counter-example #4.
- `tests/test_profile_panel.py` two test docstrings + one
  inline assertion message + one trailing comment.
- `CHANGELOG.md` Wave 2 entry.
- `ROADMAP.md` AI-Agent Track building block.

The corrected framing across all surfaces:
- Honestly state the contract: `ContinuousDiD` requires
  `P(D=0) > 0` and positive treated doses; Remark 3.1 not
  implemented.
- When the contract isn't met, say `ContinuousDiD` "as currently
  implemented does not apply" — not "do this fix."
- Mention routing alternatives that ARE in the library and DON'T
  require `P(D=0) > 0`: `HeterogeneousAdoptionDiD`, linear DiD
  with a continuous covariate. Those are routing facts, not
  methodology endorsements.
- Re-encoding stays in the prose as an "agent-side preprocessing
  choice that changes the estimand and is not documented in
  REGISTRY as a supported fallback" — explicitly NOT endorsed.

Length housekeeping: trimmed redundancy in the §4.7 trailing
paragraph (consolidated to a pointer at §2) and tightened the §2
recoding paragraph; autonomous (65984 chars) < full (66031),
`test_full_is_largest` green.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…s "negative dose" branches; HAD only valid on the former

Reviewer correctly noted that the round-15/16 wording listed
`HeterogeneousAdoptionDiD` as a routing alternative whenever
`ContinuousDiD` fails on the dose-related preflights, but HAD
itself requires non-negative dose support and raises on negative
post-period dose at `had.py:1450-1459` (paper Section 2). On a
panel with `dose_min < 0`, routing to HAD silently steers an agent
into the same fit-time error. Verified the rejection at
`had.py:1450-1459`.

Reworded every site to split the two failure modes:

- Branch (a): `has_never_treated == False` (no zero-dose controls
  but all observed doses non-negative). `ContinuousDiD` does not
  apply (Remark 3.1 not implemented). HAD IS a routing alternative
  on this branch (HAD's contract requires non-negative dose,
  satisfied here); linear DiD with a continuous covariate is
  another.
- Branch (e): `dose_min < 0` (negative treated doses).
  `ContinuousDiD` does not apply AND HAD is **not** a fallback
  either — HAD raises on negative post-period dose
  (`had.py:1450-1459`). Linear DiD with a signed continuous
  covariate is the applicable alternative on this branch.

Updated wording across:
- `diff_diff/profile.py` `TreatmentDoseShape` docstring (refactored
  from item-by-item duplication into a numbered list with a single
  "Routing alternatives when (1) or (5) fails" section that splits
  the two branches; trimmed redundancy).
- `diff_diff/guides/llms-autonomous.txt` §2 field reference (split
  the When-(1)-or-(5)-fails paragraph into the two branches).
- `diff_diff/guides/llms-autonomous.txt` §4.7 trailing paragraph
  (consolidated to a pointer at §2's split discussion).
- `diff_diff/guides/llms-autonomous.txt` §5.2 reasoning chain
  counter-example #4 (no never-treated branch: HAD applies) and
  counter-example #5 (negative-dose branch: HAD does NOT apply,
  cite `had.py:1450-1459`).
- `CHANGELOG.md` Wave 2 entry.
- `ROADMAP.md` AI-Agent Track building block.
- `tests/test_profile_panel.py` two test docstrings/comments.

Added `test_autonomous_negative_dose_path_does_not_route_to_had`
in `tests/test_guides.py` asserting that §5.2 explicitly cites
`had.py:1450-1459` on the negative-dose branch (used a single-
line fingerprint since the prose phrase "non-negative dose
support" is split across newlines in the rendered guide).

Length housekeeping: trimmed counter-example #4 and #5 prose +
§4.7 trailing paragraph to point at §2's split discussion;
autonomous (65374 chars) < full (66031), `test_full_is_largest`
green.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>