schema.json requires a reason on pairwiseDeltaAllowed, but the compare
script also runs standalone (check-bc5-compat, the scheduled workflow)
with no schema-validation step in front of it — a reason-less waiver
still flattened into ALLOW_PATHS and suppressed enforcement unaudited.
The extraction now rejects waivers without a non-empty string reason
(exit 2). Scenario R pins it; scenario P's assertion follows the
combined waiver error message.

A LIVE_RECORD_DIR (or, theoretically, a mktemp result) beginning with
'-' would be parsed as an rm option; end option parsing explicitly.

The compare + harness scripts guard for it with a clear error; the
Makefile comment now says so up front for contributors running
make check on stock macOS bash.

…waiver reason

* Each snapshot's recorded .operation must match the test entry's
  operation before comparison — a stale or overwritten file with the
  right name exits 2 instead of being compared as this test's capture.
  Scenario S pins it.

* conformance/schema.json aligns with the runtime waiver guard: the
  pairwiseDeltaAllowed conditional now sets minLength 1 on reason, so
  an empty-string reason fails schema validation the same way it fails
  the compare script. TS conformance suite re-run green (84 passed).

schema.json's minItems:1 on paths doesn't run in front of standalone
invocations (check-bc5-compat, the scheduled workflow). An empty waiver
paths array is functionally harmless (it waives nothing) but is always
a fixture mistake — reject it with the same exit-2 operator error as
the other waiver shape violations. Scenario T pins it.

… message

* The rm -rf guard now refuses '..' only as a PATH SEGMENT (leading,
  interior, trailing, or the whole path) instead of any substring, so
  benign names like tmp/live-canary..pr308 are allowed while every
  traversal form still refuses. Patterns unit-verified for all six
  shapes.

* The regression harness preflights jq with the same clear error as
  the compare script instead of failing mid-setup with a raw
  command-not-found.

* snapshot_read_error's hint no longer blames only malformed JSON —
  an invalid jq path expression built from a rule's path entry lands
  in the same handler.

The aggregate is a list, not an object — the rule flags it as an
invalid target rather than evaluating it. Say so and point rule
authors at pages[N].body.X for keys checks on paginated endpoints.

The aggregate handling keyed off any '[*]' substring, broader than the
documented pages[*].body.X form — an undocumented items[*].foo would
stream through jq with undefined comparison semantics instead of
surfacing the fixture mistake. Both the read and the supersetArray
totals branch now see only validated leading-pages[*] paths; anything
else exits 2 as an unsupported path. Scenario U pins it.

* CONTRIBUTING's prerequisites now list the shared tooling make check
  needs beyond per-SDK toolchains: jq, and bash >= 4.4 for the
  pairwise-canary scripts (macOS /bin/bash is 3.2 — brew install bash;
  the scripts fail fast with that exact hint).

* Scenario U2 pins that a second '[*]' after a valid leading pages[*]
  is rejected as unsupported (the guard already covered it).

* An absolute LIVE_RECORD_DIR equal to the repo checkout (or any of
  its ancestors, e.g. LIVE_RECORD_DIR=$PWD or $HOME) passed the
  guards and rm -rf'd the working tree before the canary ran. A new
  prefix check refuses any path the checkout lives under; absolute
  paths elsewhere (/tmp/canary) and in-checkout subdirs still work.
  Patterns unit-verified for checkout, $HOME, /tmp, relative, and
  in-checkout-absolute shapes.

* The harness's mktemp gets an explicit template — current macOS
  accepts a bare -d (empirically verified here), but older BSD
  variants insist on a template and it costs nothing.

… divergence

bc3's BC5 API train shipped 2026-07-18..21 and settled the memories question:
doc/api/sections/my_notifications.md (bc3 #11628) documents memories as an
always-empty placeholder superseded by bubble_ups, and bc3 #10947 — the
once-planned `json.memories @bubble_ups` alias — is closed unmerged. The
pairwiseDeltaAllowed entry is therefore not a temporary waiver pending a fix;
it is the permanent machine-readable record of an accepted BC4→BC5
subtractive delta.

Update the GetMyNotifications test description and both rule reasons in
live-my-surface.json, and align the canonical example + prose in
CONTRIBUTING.md, replacing the 'drop the waiver once #10947 merges' guidance
with the actual retire conditions: BC4 empties memories too, or BC5's
documented contract changes.

is_allowed's early continue ran before the '[*]' syntax guard, so an
unsupported path (items[*].foo, a double star) that also appeared in a
pairwiseDeltaAllowed waiver exited clean instead of reporting the fixture
error — a typo in a waived canary rule could permanently suppress
enforcement without ever surfacing. Validate path syntax first, then apply
the allowlist skip. Adds scenario U3 covering the waived-unsupported-path
case. Flagged by Codex review.

The unknown-RULE_TYPE rejection lived in the dispatch case inside the
per-path loop, which only runs for paths that survive the is_allowed skip.
A misspelled type (pairwiseSupersetAray) whose paths were all waived by
pairwiseDeltaAllowed therefore exited clean — an operator error silently
suppressed, and the typo'd rule never enforced anything again. Hoist the
type check to rule extraction, before any per-path processing; the case's
catch-all arm stays as defense in depth. Adds scenario V covering the
waived-misspelled-type case. Flagged by Codex review.

The star guard only matched the literal '[*]', so a bare jq-stream typo
like items[] (or pages[0].body.items[]) reached read_value unwrapped: the
stream made the length variables multi-line/empty, the -lt comparison
errored falsy, and a real shrink violation (BC4 [1,2] vs BC5 []) exited 0
— a false-green on exactly the invariant the script enforces.

Generalize the guard: brackets are only accepted as a leading pages[N] or
pages[*] segment (regex-validated), everything else is rejected as a
fixture mistake with exit 2, still ahead of the waiver skip. Adds
scenario W reproducing the items[] false-green. Flagged by Codex review.

- extract validate_rule_path() and apply it to pairwiseDeltaAllowed
  entries too: a malformed bracket path appearing only in a waiver names
  a path no enforcing rule can ever target and previously sat unnoticed
- require bracketed paths to enter the page body (pages[N].body...):
  page-level keys other than body are backend-specific metadata with no
  pairwise semantics, and a typo like pages[0].items read null on both
  snapshots and false-greened superset rules
- validate_snapshot now type-checks page keys (status number, headers
  object, bodyText string, url string) so typed corruption fails loud
  instead of reading as nulls
- check-bc5-compat rm guard: canonicalize an existing LIVE_RECORD_DIR
  (cd + pwd -P) before the checkout-ancestor check so non-canonical
  spellings can't slip past the string compare; a failed cd empties LRD,
  which the ancestor pattern then refuses

Regression scenarios X (waiver-only malformed path), Y (bracket path
skipping .body), Z (wrongly-typed page keys) added; 29/29 green.

A concrete pages[N] rule asserts the page exists on both backends. When
either capture had fewer pages, read_value returned null and the rule
passed clean — hiding a fixture typo (pages[1] against single-page
captures) or a genuine page-count regression. Both demand operator
attention: fail with exit 2 naming the short snapshot and its page
count. Cross-page-count comparison remains the job of pages[*]
aggregation.

Regression scenarios AA (absent from both) and AB (one-sided page loss,
error names the bc5 snapshot) added; 31/31 green.

…kets

A path with jq punctuation but no brackets ('items, .pages') slipped the
bracket blacklist and interpolated into the jq program as raw syntax.
Terminate the class: validate_rule_path now admits only the documented
grammar — dotted [A-Za-z0-9_-] key segments, optionally led by a
pages[N]/pages[*] indexer that must enter the page body, plus the ""
body-root sentinel — and rejects everything else as a fixture mistake.

Regression scenario AC (comma path) added; 32/32 green.

…snapshots

- validate_rule_path: remove '-' from the segment charset — '.foo-bar'
  parses in jq as subtraction ('.foo - bar'), not a key lookup, so a
  hyphen-keyed rule evaluated to garbage instead of reading the field.
  No fixture path uses hyphens; keys requiring quoting are a future,
  deliberate extension of to_jq_path. Regression scenario AD added
  (33/33 green).
- live-canary.yml: upload only the decode/ result trees. Wire snapshots
  carry raw response bodies and captured headers from a live account and
  must not persist as downloadable workflow artifacts; decode results
  carry only booleans, error strings, and dotted field paths.

The round-7 canonicalization rewrote LRD itself, so rm -rf on a
symlinked LIVE_RECORD_DIR deleted the tree the link points to instead
of the link — a behavior change with data-loss potential. Canonicalize
into a separate guard variable for the checkout-ancestor check and keep
the rm on the original (trailing-slash-stripped) path: a symlink is
removed as a link, while a symlink pointing into the checkout is still
refused by the guard (verified both by hand).

A pairwiseEqual failure printed the complete BC4 and BC5 JSON values
into the comparison output, which the scheduled canary lands in
retained, log-access-readable Actions logs — leaking live-account field
content (names, URLs, notification text) on any drift. Report only
shape summaries (type + keys/length) and point the operator at a local
re-run for raw values. Regression scenario AE proves the violation
still fires while the content stays out of the output; 34/34 green.

Completes 5270e94: the invalid-target branches of pairwiseSupersetArray
and pairwiseSupersetKeys still embedded the full BC4/BC5 JSON values in
violation output. Factor the shape summary into a shared shape_of()
helper and use it at all three sites; no violation message carries live
field content anymore (grep-verified zero BC4=$BC4_VAL sites).

Shorthand paths ("" or "foo.bar") normalize to pages[0].body — fine
for single-page captures, but against a multi-page snapshot they
silently checked only page 0 and ignored the rest, under-enforcing the
rule. Hard-error with a pointer to pages[N]/pages[*] syntax as soon as
either side's capture has more than one page. Regression scenario AF
added; 35/35 green.