Ships the two un-rowed obligations from the 3.9-cut checklist
(docs/v4-design.md section 9, items 2 and 3) - the last thing between 3.9 being
code-complete and the formal cut. Lands BEFORE the release commit, the order the
design already mandates in two normative places (the phase table at :828 and the
hand-assertion sentence at :885); the release PR verifies these rather than
shipping them.

docs/migration-4.0.md is a 21-row orientation table plus a 108-row per-symbol
appendix, worked before/after for the three merges, a codemod regex table, and a
separate "already shipped in 3.9" section. The appendix row set is derived from
the ledger by _changes_at_4_0 and asserted in both directions, so a rescheduled
or added row fails CI until the guide is updated.

THE APPENDIX HAS NO DERIVED ADVICE COLUMNS, and that is the load-bearing
decision. An `Available since` column was designed, gated, and then removed:
every label it would carry rests on a single ledger field while the fact asserted
is compound. `introduced_in` tracks the dataclass STORAGE flip, so the nine
field-flip rows say 4.0 while `.att` resolves today; a null `new` covers
removals, default flips AND translations; a successor that resolves can still
raise (the five bootstrapped-fit aggregate() families). No better probe fixes
that, so the appendix carries only what is mechanically checkable and the
compound claim moved into a hand-written Fix cell sourced from each row's notes.

That choice was validated during implementation, not assumed. The generator's
null-successor guard refused to emit a mechanical rename for M-087, then
M-011/M-012/M-014/M-016 - each needed text written from its notes. Executing the
robust= family produced THREE distinct rules across four rows: drop-or-translate
on DifferenceInDifferences/LinearRegression, drop-only on TripleDifference (it
accepts only hc1, so vcov_type="classical" RAISES), and the inverse on HAD, whose
legacy default is non-robust. One pre-written rule would have been wrong on all
four.

Local review then found three defects that prove the same point one level down,
in the hand-written prose rather than the derived table:

- The auto-cluster opt-out was INVERTED. `cluster=None` and omission are the same
  thing under the 4.0 policy - both auto-cluster at unit - so the documented
  "reproduce 3.x exactly" recipe would have silently returned clustered SEs,
  p-values and CIs. `cluster=False` disables (v4-design.md:681). Also scoped: the
  TWFE event-study mode has auto-clustered since 3.9, so the flip changes the
  static and other panel paths, not that one.
- M-139 pointed at a method that does not exist. `HADPretestReport.aggregate` is
  a metadata field; the successor is panel-shape inference. The advice raised.
- All three merge examples carried an invalid keyword: `treated=` for
  `treatment=`, `partition=` for `eligibility=` on the RETIRED staggered class
  (the surviving TripleDifference does take `partition=`, which is what made it
  look right), and `treated=`/`post=` for `treatment=`/`time=` on both CiC sides.

The third is the failure mode this page is structurally exposed to - it is
markdown, and test_doc_snippets.py discovers only RST, so a plausible-but-wrong
keyword ships silently. Rather than leave that acceptance bare,
test_migration_guide_examples_bind_to_real_signatures ast-parses the guide's
python blocks and asserts every constructor/fit() keyword exists on the target
signature. It binds examples to reality WITHOUT executing 4.0-only behaviour,
which is why the snippet harness could not do this job. Verified non-vacuous
rather than assumed: replayed against the four original broken keywords it flags
all four, and it checks 33 keywords on the live guide.

The R argument table is scoped to the three packages r_comparison.rst actually
evidences with paired R/Python blocks - did, HonestDiD, synthdid. fixest,
DIDmultiplegtDYN and DIDHAD appear there in prose only, so they are named as out
of scope rather than mapped from memory. Non-1:1 cases get a Notes cell instead
of a fabricated row: aggte(type="dynamic") -> aggregate(type="event_study")
(AGGREGATION_VOCABULARY has no "dynamic"), synthdid's block structure ->
treatment indicator plus post_periods, and HonestDiD's betahat/sigma carried by
the results object. Its gate is two-directional and resolves each name against
the SPECIFIC callable the row names - resolving against any public callable would
let `time` or `unit` be satisfied by an unrelated signature. The R side itself is
ungateable (no R signatures in the repo) and the docstring says so. The gate
earned itself immediately by catching a real defect in the table it guards: a
cell read `aggregate(type=)` where the parameter is `type`.

Two deviations from the approved plan, both forced by execution. The doc-deps
union needed a private-module filter the plan did not specify - the mechanical
recipe resolved `_reporting_helpers.py` to a key that does not exist; excluding
private modules loses no coverage (both public owners are already in the union)
and avoids a schema edit. And the no-op Fix denylist had to become word-boundary
regex: a bare "none" fired on the legitimate `cluster=None` and
`summary(alpha=None)` that two cells must state.

TODO.md's aggregate= row is re-graded Quick -> Heavy with a measured inventory:
28 executable code-cell sites across 9 notebooks (14_continuous_did and
15_efficient_did match only in markdown prose; 21_had_pretest_workflow uses a
spaced `aggregate =` that a naive grep misses), plus six docs files of which two
API pages cannot migrate at all today because their report consumers read the raw
event_study_effects field post-fit aggregate() never populates.

Also corrects two false statements on the R page that no snippet gate can reach:
`.ci` (the field is conf_int) and the claim that aggregation is requested at fit
time.

The llms.txt and README.md catalog entries are deliberately NOT here: both list
pages as readthedocs.io/en/stable/..., and `stable` builds the newest tag
(v3.8.0), so an entry added before the cut is a 404 in the wheel-bundled agent
contract. They go with the release commit.

Cross-references to TODO.md/DEFERRED.md/v4-design.md/the ledger YAML are inline
literals, never links - none is a Sphinx document, so a MyST link to any of them
fails the -W build.

Verified: 469 gate tests pass (matrix incl. 8 new guide gates + 3 parser vacuity
tests, docs IA incl. the new R-table gate, doc-deps integrity); naming guard
clean (the guide is outside every scan lane); doc snippets 119 passed / 5 skipped
with r_comparison block indices unshifted; Sphinx -W build succeeded with zero
warnings; black, ruff, and mypy clean at the CI-pinned toolchain.

…arry their values

Addresses the CI review on #759 (1 P1 + 3 P2, all verified against source before
acting).

P1 - THE SYNTHETICDID ADVICE WOULD HAVE CHANGED ESTIMATES. The appendix told
readers to drop `lambda_reg`/`zeta` and "set `zeta_omega=`/`zeta_lambda=`
instead". But those params have been IGNORED since 3.0.0 - `synthetic_did.py:216`
warns "deprecated and ignored. Regularization is now auto-computed from data. Use
zeta_omega to OVERRIDE". So 3.x behaviour is auto-computed regularization, and
copying the old value into the new parameter activates an override that was
previously inert, changing unit weights, ATT and inference. The value-preserving
migration is to delete the argument and leave the new ones unset. Both cells now
say that explicitly and warn against the mechanical copy. The ledger's own "use
zeta_omega" note describes where the CAPABILITY moved, not a value migration -
reading it as the latter is what produced the error.

P2 - the results-field section documented only `overall_att` -> `att`, but the
whole inference quintet moves (`overall_se`, `overall_t_stat`, `overall_p_value`,
`overall_conf_int`). Added the mapping table, plus the ContinuousDiD outlier that
M-058's notes call out: its siblings carry an `att` infix
(`overall_att_se`, `overall_att_t_stat`, ...), so anyone grepping `overall_se`
misses them entirely.

P2 - the codemod table claimed to cover identifier renames but omitted
`treatment_col` -> `takeup` (the one `*_col` param whose replacement is not just
the suffix dropped), `aggregation` -> `level`, and `controls` -> `covariates`.
Added, and the three renames that CANNOT be global replacements are now
enumerated with the reason each is contextual: `group=` -> `unit=` would corrupt
`TripleDifference.fit`, which takes a `group=` that is not renamed; `time=` ->
`post=` survives everywhere else as the calendar column; `aggregate=` is not a
rename at all.

P2 - the signature-binding gate added in the previous round was itself vacuous:
deleting every example, or renaming the ```python fence, emptied its loop and it
passed. It now asserts a floor of 4 blocks and 20 resolvable keywords, and names
the six estimators whose worked examples must be present. Verified by mutation
rather than assumed - renaming the fence and deleting the QDiD example each fire
the expected assertion, and the guide is restored afterwards.

The gate I wrote to catch documentation errors had the same defect I built three
parser-vacuity tests to prevent on the table parsers. Adding a floor to a new
gate is not optional just because the gate is small.

Verified: 469 gate tests pass; Sphinx -W build succeeded with zero warnings;
black and ruff clean.