<?xml version="1.0" encoding="utf-8"?>
<entry xmlns="http://www.w3.org/2005/Atom" xml:base="https://docs.spec-kitty.ai/adr/3.x/2026-09-14-1-census-floor-ratchet-adjudication/">
  <id>https://docs.spec-kitty.ai/adr/3.x/2026-09-14-1-census-floor-ratchet-adjudication/</id>
  <title>ADR: Census-Floor Ratchets — Per-Ratchet Adjudication (Tripwire, Retire, Retire, Keep-Property)</title>
  <updated>2026-09-14T00:00:00.000Z</updated>
  <author><name>Spec Kitty</name></author>
  <link rel="alternate" type="text/html" href="https://docs.spec-kitty.ai/adr/3.x/2026-09-14-1-census-floor-ratchet-adjudication/"/>
  <summary>Per-ratchet verdicts on the #4315 census-floor ratchets: the routed-meta floor is deleted outright, golden-count and shard markers retire, property gates stay.</summary>
  <content type="html">&lt;h2 id=&quot;context-and-problem-statement&quot;&gt;Context and Problem Statement&lt;/h2&gt;
&lt;p&gt;Issue &lt;a href=&quot;https://github.com/spec-kitty/spec-kitty/issues/4315&quot;&gt;#4315&lt;/a&gt; (parent epic
&lt;a href=&quot;https://github.com/spec-kitty/spec-kitty/issues/1931&quot;&gt;#1931&lt;/a&gt;) asks whether a
&lt;strong&gt;census count&lt;/strong&gt; is the right instrument for four architectural gates that
demand a manual baseline re-pin on ordinary changes:&lt;/p&gt;
&lt;ol&gt;
&lt;li&gt;&lt;code dir=&quot;auto&quot;&gt;tests/architectural/test_inline_meta_read_gate.py::test_routed_load_meta_floor&lt;/code&gt;
(&lt;code dir=&quot;auto&quot;&gt;ROUTED_LOAD_META_FLOOR&lt;/code&gt;);&lt;/li&gt;
&lt;li&gt;&lt;code dir=&quot;auto&quot;&gt;tests/architectural/test_golden_count_ban.py&lt;/code&gt; (&lt;code dir=&quot;auto&quot;&gt;_golden_count_baseline.json&lt;/code&gt;);&lt;/li&gt;
&lt;li&gt;&lt;code dir=&quot;auto&quot;&gt;tests/architectural/marker_baseline.txt&lt;/code&gt;;&lt;/li&gt;
&lt;li&gt;the &lt;code dir=&quot;auto&quot;&gt;tests/_arch_shard_map.py&lt;/code&gt; / &lt;code dir=&quot;auto&quot;&gt;tests/_next_shard_map.py&lt;/code&gt; completeness counts.&lt;/li&gt;
&lt;/ol&gt;
&lt;p&gt;The trigger was the PR &lt;a href=&quot;https://github.com/spec-kitty/spec-kitty/pull/4257&quot;&gt;#4257&lt;/a&gt;
landing pass: a ~277-line single-feature change tripped two of them. The routed
floor went red because a new function, &lt;code dir=&quot;auto&quot;&gt;record_discard&lt;/code&gt;, read &lt;code dir=&quot;auto&quot;&gt;meta.json&lt;/code&gt;
&lt;strong&gt;through the canonical &lt;code dir=&quot;auto&quot;&gt;_require_meta&lt;/code&gt; seam&lt;/strong&gt; — precisely the behaviour the gate
exists to reward. As #4315 puts it: &lt;em&gt;“The gate punished the behavior it exists to
encourage.”&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;#4315 requires a &lt;strong&gt;per-ratchet&lt;/strong&gt; decision — keep-hard / demote-advisory /
replace-with-property-check / auto-maintain / retire — &lt;em&gt;“Not a blanket removal
without that adjudication.”&lt;/em&gt; This ADR is that adjudication.&lt;/p&gt;
&lt;h3 id=&quot;two-scope-corrections&quot;&gt;Two scope corrections&lt;/h3&gt;
&lt;p&gt;Measured against the tree at &lt;code dir=&quot;auto&quot;&gt;2222506cc5&lt;/code&gt;, two of #4315’s four items are not what
the ticket assumed:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;&lt;code dir=&quot;auto&quot;&gt;marker_baseline.txt&lt;/code&gt; does not exist.&lt;/strong&gt; &lt;code dir=&quot;auto&quot;&gt;tests/architectural/test_marker_baseline.py&lt;/code&gt;
was deleted by &lt;code dir=&quot;auto&quot;&gt;177e062694&lt;/code&gt; (&lt;em&gt;“test: assertively sanitize low-signal suite
cruft”&lt;/em&gt;, &lt;a href=&quot;https://github.com/spec-kitty/spec-kitty/issues/3285&quot;&gt;#3285&lt;/a&gt;,
2026-08-12); only a stale &lt;code dir=&quot;auto&quot;&gt;.pyc&lt;/code&gt; remains. It is already retired.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The shard-map “completeness counts” are not counts.&lt;/strong&gt; Both live tests are
set-partition properties over a freshly collected corpus. There is no pinned
integer anywhere in the shard machinery. The friction is the hand-maintained
&lt;strong&gt;assignment table&lt;/strong&gt;, and it is confined to one of the two groups.&lt;/li&gt;
&lt;/ul&gt;
&lt;h3 id=&quot;the-governing-constraint-what-doctrine-actually-requires&quot;&gt;The governing constraint: what doctrine actually requires&lt;/h3&gt;
&lt;p&gt;&lt;code dir=&quot;auto&quot;&gt;DIRECTIVE_043&lt;/code&gt; &lt;em&gt;Close Defect Classes by Construction&lt;/em&gt; carries
&lt;code dir=&quot;auto&quot;&gt;enforcement: required&lt;/code&gt;, and its integrity rules bind here:&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;A gate that trivially passes when the relevant call-site count is zero
(vacuous gate) is non-compliant — &lt;strong&gt;the gate must have a concrete floor&lt;/strong&gt;.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;The companion &lt;code dir=&quot;auto&quot;&gt;tactic:architectural-gate-non-vacuity&lt;/code&gt; states the floor’s purpose
precisely, and its own worked examples set the scale:&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;A route-or-allowlist gate is &lt;strong&gt;vacuously satisfiable&lt;/strong&gt; if all existing calls are
allowlisted and the count through the sanctioned surface &lt;strong&gt;drops to zero&lt;/strong&gt;.
Assert separately that the count of calls through the canonical (routed) surface
is above a minimum.&lt;/p&gt;
&lt;p&gt;— examples: &lt;code dir=&quot;auto&quot;&gt;assert canonical_call_count &gt;= 1&lt;/code&gt; · &lt;code dir=&quot;auto&quot;&gt;assert len(...) &gt;= MIN_ROUTED_FILES&lt;/code&gt;&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;&lt;strong&gt;Nothing in doctrine requires a &lt;em&gt;separate routed-count floor&lt;/em&gt; on this gate at
all.&lt;/strong&gt; DIRECTIVE_043’s mandate is that the gate not pass vacuously; for the
inline-read gate that guarantee is already carried by mechanic 2 — the ceiling’s
&lt;code dir=&quot;auto&quot;&gt;MARGIN&lt;/code&gt; clause (&lt;code dir=&quot;auto&quot;&gt;INLINE_META_READ_FLOOR - live &amp;#x3C;= MARGIN&lt;/code&gt;, so the live count must
stay near the ceiling and a broken or empty scan reds) — and by the FR-010
single-decoder gate. The routed-count floor measured routing &lt;em&gt;through&lt;/em&gt; the
canonical reader — a reward signal, not the gate’s non-vacuity guard — and a
tighter, floor-independent de-routing detector already exists (&lt;code dir=&quot;auto&quot;&gt;_ACCOUNTED_SITES&lt;/code&gt;
in &lt;code dir=&quot;auto&quot;&gt;test_meta_fail_closed_full_census_contract.py&lt;/code&gt;, exact-equality, keyed on
&lt;code dir=&quot;auto&quot;&gt;load_meta_fail_closed&lt;/code&gt;). The routed floor therefore added no non-vacuity
protection the gate did not already have; deleting it leaves DIRECTIVE_043 and the
tactic fully satisfied. This ADR records the point explicitly so no future
reviewer reads the deletion below as weakening DIRECTIVE_043.&lt;/p&gt;
&lt;p&gt;Two further facts bound how much authority the current shape carries:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;The tactic’s own &lt;code dir=&quot;auto&quot;&gt;notes&lt;/code&gt; audit (T007) labels &lt;strong&gt;three of its four elements
ASPIRATIONAL&lt;/strong&gt;, including the concrete floor and the routed-count floor: &lt;em&gt;“it is
aspirational relative to the current exemplar.”&lt;/em&gt;&lt;/li&gt;
&lt;li&gt;&lt;code dir=&quot;auto&quot;&gt;tactic:frozen-baseline-shrink-only-ratchet&lt;/code&gt; — the canonical maintenance
discipline — governs &lt;strong&gt;debt ceilings&lt;/strong&gt;: &lt;em&gt;“CI fails if the live count &lt;strong&gt;grows&lt;/strong&gt;
beyond the committed baseline, and warns (but does not fail) if it shrinks,”&lt;/em&gt;
with a burn-down target of zero. Every entry in the canonical
&lt;code dir=&quot;auto&quot;&gt;tests/architectural/_baselines.yaml&lt;/code&gt; has that shape: a count of &lt;em&gt;bad&lt;/em&gt; things
that must shrink.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;That distinction is the analytical core of this ADR. A &lt;strong&gt;debt ceiling&lt;/strong&gt; has a
natural fixed point (zero) and stabilises as the debt drains. A &lt;strong&gt;goodness floor&lt;/strong&gt;
— a minimum count of &lt;em&gt;good&lt;/em&gt; call sites, pinned within a margin of live — has no
fixed point: good call sites grow with the codebase, so the pin must be chased
forever. The doctrine sanctions the former. The latter is an invention of one
gate.&lt;/p&gt;
&lt;h2 id=&quot;decision-drivers&quot;&gt;Decision Drivers&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Measured catch record vs. measured landing friction&lt;/strong&gt;, per gate — not opinion.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Doctrinal compliance&lt;/strong&gt;: DIRECTIVE_043 is &lt;code dir=&quot;auto&quot;&gt;required&lt;/code&gt;; no verdict may leave a
route-or-allowlist gate vacuously satisfiable.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Single canonical authority&lt;/strong&gt; (charter governing principle): one baseline
authority, not two.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The #2913 operator ruling&lt;/strong&gt; binds this class of change: &lt;em&gt;“because the outcome
may alter or remove a blocking, shared CI gate — land the CaaCS tally and your
recommendation first, and get maintainer concurrence on the chosen decision-menu
option before implementing it.”&lt;/em&gt; Concurrence for every verdict below was
obtained 2026-09-14 before any gate change was authored.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id=&quot;evidence&quot;&gt;Evidence&lt;/h2&gt;
&lt;p&gt;All figures measured against &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt; @ &lt;code dir=&quot;auto&quot;&gt;2222506cc5&lt;/code&gt; on 2026-09-14.&lt;/p&gt;
&lt;h3 id=&quot;gate-1--routed_load_meta_floor&quot;&gt;Gate 1 — &lt;code dir=&quot;auto&quot;&gt;ROUTED_LOAD_META_FLOOR&lt;/code&gt;&lt;/h3&gt;






























&lt;table&gt;&lt;thead&gt;&lt;tr&gt;&lt;th&gt;Bucket&lt;/th&gt;&lt;th&gt;Count&lt;/th&gt;&lt;th&gt;Meaning&lt;/th&gt;&lt;/tr&gt;&lt;/thead&gt;&lt;tbody&gt;&lt;tr&gt;&lt;td&gt;A. False red&lt;/td&gt;&lt;td&gt;2 (self-reported as &lt;em&gt;“the third recurrence”&lt;/em&gt;)&lt;/td&gt;&lt;td&gt;Routing coverage did not regress; the census moved on a delegation/dedup/rename shape change&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;B. Drift re-pin&lt;/td&gt;&lt;td&gt;6&lt;/td&gt;&lt;td&gt;Genuine drift accumulated by unrelated merges; the landing PR added no routed sites&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;C. Legitimate own-diff re-pin&lt;/td&gt;&lt;td&gt;15&lt;/td&gt;&lt;td&gt;The PR correctly added routed sites — the gate demanded manual work for doing the right thing&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;&lt;strong&gt;D. Real catch&lt;/strong&gt;&lt;/td&gt;&lt;td&gt;&lt;strong&gt;0&lt;/strong&gt;&lt;/td&gt;&lt;td&gt;—&lt;/td&gt;&lt;/tr&gt;&lt;/tbody&gt;&lt;/table&gt;
&lt;p&gt;&lt;strong&gt;23 value-moving commits in 63 days&lt;/strong&gt; (2026-07-08 → 2026-09-09). Bucket D is
empty: the floor has never once blocked a genuine de-routing.&lt;/p&gt;
&lt;p&gt;Three further facts decide this gate:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Zero headroom today.&lt;/strong&gt; Live routed census = &lt;strong&gt;157&lt;/strong&gt;, floor = &lt;strong&gt;153&lt;/strong&gt;, margin =
&lt;strong&gt;4&lt;/strong&gt;. The permitted band is &lt;code dir=&quot;auto&quot;&gt;154 ≤ live ≤ 157&lt;/code&gt;; the repo sits on its upper
edge. The next PR anywhere that adds one canonical &lt;code dir=&quot;auto&quot;&gt;load_meta*&lt;/code&gt; call reds the
gate.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The attack the floor deters was never attempted.&lt;/strong&gt;
&lt;code dir=&quot;auto&quot;&gt;inline_meta_read_allowlist.yaml&lt;/code&gt; has &lt;strong&gt;3 commits ever&lt;/strong&gt; — creation, one
qualname freshening, one path rewrite for the &lt;code dir=&quot;auto&quot;&gt;src/charter/&lt;/code&gt; move — and &lt;strong&gt;net
zero entries added in 68 days&lt;/strong&gt;. &lt;code dir=&quot;auto&quot;&gt;INLINE_META_READ_FLOOR&lt;/code&gt; has never moved off
&lt;strong&gt;7&lt;/strong&gt;. Mass-allow-listing is in any case visible in review as N new
&lt;code dir=&quot;auto&quot;&gt;{key, rationale, issue}&lt;/code&gt; entries &lt;em&gt;plus&lt;/em&gt; a ceiling raise.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The one real defect this file ever caught belongs to the ceiling, not the
floor.&lt;/strong&gt; &lt;code dir=&quot;auto&quot;&gt;0df0dbd59e&lt;/code&gt; (&lt;em&gt;“route cut-over meta reads through load_meta (BOM
fail-closed inversion)”&lt;/em&gt;) drained two hand-rolled &lt;code dir=&quot;auto&quot;&gt;json.loads&lt;/code&gt; reads — a
fail-closed verdict inverted in the wrong direction. That commit is explicit
about which half fired: &lt;em&gt;“&lt;code dir=&quot;auto&quot;&gt;INLINE_META_READ_FLOOR&lt;/code&gt; is deliberately NOT bumped:
draining these two reads put the inline census back under the existing ceiling
on its own.”&lt;/em&gt; The routed floor was re-pinned only as collateral.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;Cost: &lt;strong&gt;1,551 exclusive lines&lt;/strong&gt;, of which ~100 are a hand-written dated re-pin
ledger inside one constant’s comment block. Of 32 commits touching the file, 9 are
pure re-pins, 5 are genuine logic changes, and 18 are unrelated changes forced to
carry a floor bump.&lt;/p&gt;
&lt;h3 id=&quot;gate-2--test_golden_count_ban&quot;&gt;Gate 2 — &lt;code dir=&quot;auto&quot;&gt;test_golden_count_ban&lt;/code&gt;&lt;/h3&gt;
&lt;p&gt;Already demoted to advisory by
&lt;a href=&quot;https://github.com/spec-kitty/spec-kitty/issues/3458&quot;&gt;#3458&lt;/a&gt; / mission
&lt;code dir=&quot;auto&quot;&gt;ci-pipeline-reinstatement-01M1X35E&lt;/code&gt; WP16, on this recorded evidence:&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;a single benign symbol addition anywhere in &lt;code dir=&quot;auto&quot;&gt;tests/&lt;/code&gt; was forcing either a
spurious escape-hatch annotation or a whole-tree re-freeze on every unrelated
PR, &lt;strong&gt;for zero real catches&lt;/strong&gt; (#3458’s own evidence: &lt;strong&gt;0 catches, 2 forced
annotations&lt;/strong&gt;).&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;Since the demotion (2026-09-07 → 2026-09-14): &lt;strong&gt;zero&lt;/strong&gt; conversions, &lt;strong&gt;zero&lt;/strong&gt;
baseline actions — and &lt;strong&gt;three PRs still paid the annotation toll&lt;/strong&gt;, the most
recent &lt;code dir=&quot;auto&quot;&gt;71d4282521&lt;/code&gt; (#4306) on 2026-09-14. Contributors keep paying a charge the
gate no longer levies.&lt;/p&gt;
&lt;p&gt;Lifetime baseline history: 15 commits — &lt;strong&gt;13 whole-tree re-freezes forced by
unrelated additions&lt;/strong&gt;, 2 genuine burn-downs, &lt;strong&gt;0 real catches&lt;/strong&gt;. The escape-hatch
toll is larger still: &lt;strong&gt;127 commits&lt;/strong&gt; adding annotations, &lt;strong&gt;387 live annotation
sites across 194 files&lt;/strong&gt;. The classifier also needed a correctness fix of its own
(&lt;code dir=&quot;auto&quot;&gt;f54c14e945&lt;/code&gt;, &lt;em&gt;“stop golden-count gate taxing honest dynamic-result cardinality
asserts”&lt;/em&gt;) — it was demonstrably taxing correct code.&lt;/p&gt;
&lt;p&gt;It is currently &lt;strong&gt;stale-advisory in five directories&lt;/strong&gt; untouched by any recent
work: &lt;code dir=&quot;auto&quot;&gt;tests/auth&lt;/code&gt; +1, &lt;code dir=&quot;auto&quot;&gt;tests/ci&lt;/code&gt; +1, &lt;code dir=&quot;auto&quot;&gt;tests/doctrine&lt;/code&gt; +4, &lt;code dir=&quot;auto&quot;&gt;tests/retrospective&lt;/code&gt;
24 vs 21, &lt;code dir=&quot;auto&quot;&gt;tests/specify_cli&lt;/code&gt; 240 vs 238. Nobody has acted on the warning.&lt;/p&gt;
&lt;p&gt;Cost: &lt;strong&gt;993 exclusive lines&lt;/strong&gt;. &lt;code dir=&quot;auto&quot;&gt;_golden_count_baseline.json&lt;/code&gt; is also a &lt;strong&gt;second
baseline authority&lt;/strong&gt; outside the canonical &lt;code dir=&quot;auto&quot;&gt;tests/architectural/_baselines.yaml&lt;/code&gt;,
against the charter’s single-canonical-authority principle.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Correction (post-spec adversarial squad, 2026-09-14).&lt;/strong&gt; An earlier draft of this
ADR counted &lt;code dir=&quot;auto&quot;&gt;test_shape_guard_membership.py&lt;/code&gt; (191 lines) as machinery that “exists
only to police the demotion”. That is &lt;strong&gt;false&lt;/strong&gt;, and the error would have caused a
real enforcement loss. Of its six tests only
&lt;code dir=&quot;auto&quot;&gt;test_shape_guard_demotion_does_not_raise_under_a_manufactured_breach&lt;/code&gt; is
golden-count-specific. Two others hold the &lt;strong&gt;C-007 enforcement canon&lt;/strong&gt;:
&lt;code dir=&quot;auto&quot;&gt;test_enforcement_allowlist_set_is_exactly_the_c007_canon&lt;/code&gt; pins the four
always-on allowlists by set-equality against a tuple held in test &lt;em&gt;code&lt;/em&gt; (so a
yaml-only edit can neither demote a member out nor promote one in), and
&lt;code dir=&quot;auto&quot;&gt;test_enforcement_allowlists_still_carry_real_blocking_assertions&lt;/code&gt; AST-proves each
still contains a real &lt;code dir=&quot;auto&quot;&gt;assert&lt;/code&gt; inside a &lt;code dir=&quot;auto&quot;&gt;test_*&lt;/code&gt; function — an anti-silent-demotion
guard. Its subjects are exactly the four gates this ADR promises to leave
untouched.&lt;/p&gt;
&lt;p&gt;Two further facts make the loss concrete. &lt;code dir=&quot;auto&quot;&gt;test_p1_planted_regression.py:242&lt;/code&gt;
branches on &lt;code dir=&quot;auto&quot;&gt;_MEMBERSHIP_PATH.exists()&lt;/code&gt; and &lt;strong&gt;returns early&lt;/strong&gt; on a substring check,
so deleting the governing test while keeping the yaml would degrade P1’s T080 to
“these three names appear somewhere in a YAML file no test governs”. And its
&lt;code dir=&quot;auto&quot;&gt;_ENFORCEMENT_GATE_NAMES&lt;/code&gt; lists only &lt;strong&gt;three&lt;/strong&gt; gates — &lt;code dir=&quot;auto&quot;&gt;test_integration_boundary.py&lt;/code&gt;
is in the yaml canon but absent from that fallback, so the membership test is its
only machine cover.&lt;/p&gt;
&lt;h3 id=&quot;gate-3--marker-baseline&quot;&gt;Gate 3 — marker baseline&lt;/h3&gt;
&lt;p&gt;Already retired by &lt;code dir=&quot;auto&quot;&gt;177e062694&lt;/code&gt; (#3285). No action.&lt;/p&gt;
&lt;h3 id=&quot;gate-4--shard-map-completeness&quot;&gt;Gate 4 — shard-map completeness&lt;/h3&gt;
&lt;p&gt;The two live tests are &lt;strong&gt;properties, not counts&lt;/strong&gt;:&lt;/p&gt;
&lt;div class=&quot;expressive-code&quot;&gt;&lt;figure class=&quot;frame not-content&quot;&gt;&lt;figcaption class=&quot;header&quot;&gt;&lt;/figcaption&gt;&lt;pre data-language=&quot;python&quot;&gt;&lt;code&gt;&lt;div class=&quot;ec-line&quot;&gt;&lt;div class=&quot;code&quot;&gt;&lt;span style=&quot;--0:#919F9F;--1:#5F636F&quot;&gt;# test_arch_shard_marker_completeness.py:44&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div class=&quot;ec-line&quot;&gt;&lt;div class=&quot;code&quot;&gt;&lt;span style=&quot;--0:#D6DEEB;--1:#403F53&quot;&gt;invalid &lt;/span&gt;&lt;span style=&quot;--0:#C792EA;--1:#8844AE&quot;&gt;=&lt;/span&gt;&lt;span style=&quot;--0:#D6DEEB;--1:#403F53&quot;&gt; {r[&lt;/span&gt;&lt;span style=&quot;--0:#D9F5DD;--1:#111111&quot;&gt;&quot;&lt;/span&gt;&lt;span style=&quot;--0:#ECC48D;--1:#984E4D&quot;&gt;nodeid&lt;/span&gt;&lt;span style=&quot;--0:#D9F5DD;--1:#111111&quot;&gt;&quot;&lt;/span&gt;&lt;span style=&quot;--0:#D6DEEB;--1:#403F53&quot;&gt;]: &lt;/span&gt;&lt;span style=&quot;--0:#82AAFF;--1:#3B61B0&quot;&gt;...&lt;/span&gt;&lt;span style=&quot;--0:#D6DEEB;--1:#403F53&quot;&gt; &lt;/span&gt;&lt;span style=&quot;--0:#C792EA;--1:#8844AE&quot;&gt;for&lt;/span&gt;&lt;span style=&quot;--0:#D6DEEB;--1:#403F53&quot;&gt; r &lt;/span&gt;&lt;span style=&quot;--0:#C792EA;--1:#8844AE&quot;&gt;in&lt;/span&gt;&lt;span style=&quot;--0:#D6DEEB;--1:#403F53&quot;&gt; records &lt;/span&gt;&lt;span style=&quot;--0:#C792EA;--1:#8844AE&quot;&gt;if&lt;/span&gt;&lt;span style=&quot;--0:#D6DEEB;--1:#403F53&quot;&gt; &lt;/span&gt;&lt;span style=&quot;--0:#C5E478;--1:#3B61B0&quot;&gt;len&lt;/span&gt;&lt;span style=&quot;--0:#D6DEEB;--1:#403F53&quot;&gt;(&lt;/span&gt;&lt;span style=&quot;--0:#C5E478;--1:#3B61B0&quot;&gt;set&lt;/span&gt;&lt;span style=&quot;--0:#D6DEEB;--1:#403F53&quot;&gt;(&lt;/span&gt;&lt;span style=&quot;--0:#82AAFF;--1:#3B61B0&quot;&gt;r&lt;/span&gt;&lt;span style=&quot;--0:#D6DEEB;--1:#403F53&quot;&gt;[&lt;/span&gt;&lt;span style=&quot;--0:#D9F5DD;--1:#111111&quot;&gt;&quot;&lt;/span&gt;&lt;span style=&quot;--0:#ECC48D;--1:#984E4D&quot;&gt;markers&lt;/span&gt;&lt;span style=&quot;--0:#D9F5DD;--1:#111111&quot;&gt;&quot;&lt;/span&gt;&lt;span style=&quot;--0:#D6DEEB;--1:#403F53&quot;&gt;])&lt;/span&gt;&lt;span style=&quot;--0:#82AAFF;--1:#3B61B0&quot;&gt; &lt;/span&gt;&lt;span style=&quot;--0:#C792EA;--1:#8844AE&quot;&gt;&amp;#x26;&lt;/span&gt;&lt;span style=&quot;--0:#82AAFF;--1:#3B61B0&quot;&gt; markers&lt;/span&gt;&lt;span style=&quot;--0:#D6DEEB;--1:#403F53&quot;&gt;) &lt;/span&gt;&lt;span style=&quot;--0:#C792EA;--1:#8844AE&quot;&gt;!=&lt;/span&gt;&lt;span style=&quot;--0:#D6DEEB;--1:#403F53&quot;&gt; &lt;/span&gt;&lt;span style=&quot;--0:#F78C6C;--1:#AA0982&quot;&gt;1&lt;/span&gt;&lt;span style=&quot;--0:#D6DEEB;--1:#403F53&quot;&gt;}&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div class=&quot;ec-line&quot;&gt;&lt;div class=&quot;code&quot;&gt;&lt;span style=&quot;--0:#C792EA;--1:#8844AE&quot;&gt;assert&lt;/span&gt;&lt;span style=&quot;--0:#D6DEEB;--1:#403F53&quot;&gt; &lt;/span&gt;&lt;span style=&quot;--0:#C792EA;--1:#8844AE&quot;&gt;not&lt;/span&gt;&lt;span style=&quot;--0:#D6DEEB;--1:#403F53&quot;&gt; invalid   &lt;/span&gt;&lt;span style=&quot;--0:#919F9F;--1:#5F636F&quot;&gt;# every collected node carries exactly one shard marker&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/figure&gt;&lt;/div&gt;
&lt;p&gt;&lt;code dir=&quot;auto&quot;&gt;test_fast_tier_marker_completeness.py&lt;/code&gt; has the same shape and parses its roots
and vocabulary &lt;strong&gt;live out of the &lt;code dir=&quot;auto&quot;&gt;Makefile&lt;/code&gt;&lt;/strong&gt;, never a hand-copied list.&lt;/p&gt;
&lt;p&gt;The friction is the hand-maintained assignment table, and it is asymmetric.
&lt;code dir=&quot;auto&quot;&gt;arch&lt;/code&gt; sets &lt;code dir=&quot;auto&quot;&gt;default_fallback=True&lt;/code&gt; (&lt;code dir=&quot;auto&quot;&gt;tests/_arch_shard_map.py:370&lt;/code&gt;), so an
unregistered file gets a deterministic hash bucket and needs no edit; &lt;strong&gt;120 of its
181 files (66%) ride the fallback with zero maintenance commits&lt;/strong&gt;. &lt;code dir=&quot;auto&quot;&gt;next&lt;/code&gt;
deliberately did not opt in, and generated &lt;strong&gt;25 of the last 28&lt;/strong&gt; maintenance
commits.&lt;/p&gt;
&lt;p&gt;But the decisive fact is that the markers select nothing. &lt;code dir=&quot;auto&quot;&gt;pytest.ini&lt;/code&gt; says so in
its own marker help text, for all six:&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;code dir=&quot;auto&quot;&gt;arch_shard_1: … (history: minted for the arch-adversarial CI pole … **that pole no longer exists and no live CI job selects the shard**)&lt;/code&gt;&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;Verified independently: &lt;code dir=&quot;auto&quot;&gt;arch_shard_*&lt;/code&gt; / &lt;code dir=&quot;auto&quot;&gt;next_shard_*&lt;/code&gt; appear in no workflow, no
&lt;code dir=&quot;auto&quot;&gt;Makefile&lt;/code&gt; target, and no script. Live CI sharding is a &lt;strong&gt;different, independent&lt;/strong&gt;
mechanism — &lt;code dir=&quot;auto&quot;&gt;.github/ci-module-registry.yml&lt;/code&gt; → &lt;code dir=&quot;auto&quot;&gt;ci-modules.yml&lt;/code&gt; expands 17 module
rows into 34 matrix leaves, selecting by &lt;strong&gt;directory&lt;/strong&gt;, not by marker.&lt;/p&gt;
&lt;p&gt;The defect class was real until &lt;code dir=&quot;auto&quot;&gt;e8cc2f444f&lt;/code&gt; (2026-08-27) deleted the
pre-programme workflows that ran these markers as matrix legs. Every registration
commit before that date was a genuine catch; every one after is pure bookkeeping.&lt;/p&gt;
&lt;p&gt;The gate cannot even keep its own table healthy: &lt;code dir=&quot;auto&quot;&gt;tests/_next_shard_map.py&lt;/code&gt; holds
&lt;strong&gt;two duplicate rows&lt;/strong&gt; — &lt;code dir=&quot;auto&quot;&gt;tests/runtime/test_upgrade_preview_bootstrap.py&lt;/code&gt; at
&lt;code dir=&quot;auto&quot;&gt;:114&lt;/code&gt; and &lt;code dir=&quot;auto&quot;&gt;:126&lt;/code&gt;, &lt;code dir=&quot;auto&quot;&gt;tests/runtime/next/test_cli_guard_family.py&lt;/code&gt; at &lt;code dir=&quot;auto&quot;&gt;:98&lt;/code&gt; and
&lt;code dir=&quot;auto&quot;&gt;:131&lt;/code&gt; — which &lt;code dir=&quot;auto&quot;&gt;dict.fromkeys&lt;/code&gt; silently resolves to shard 2. The completeness test
inspects &lt;em&gt;applied markers&lt;/em&gt; (always exactly one, by construction of the merged
dict), so it is structurally blind to the corruption.&lt;/p&gt;
&lt;h2 id=&quot;decision-outcome&quot;&gt;Decision Outcome&lt;/h2&gt;
&lt;p&gt;&lt;strong&gt;Per-ratchet, not blanket.&lt;/strong&gt; Each verdict below carries maintainer concurrence
obtained 2026-09-14 per the #2913 ruling.&lt;/p&gt;
&lt;h3 id=&quot;1-routed_load_meta_floor--delete-the-count-floor-outright&quot;&gt;1. &lt;code dir=&quot;auto&quot;&gt;ROUTED_LOAD_META_FLOOR&lt;/code&gt; → &lt;strong&gt;delete the count-floor outright&lt;/strong&gt;&lt;/h3&gt;
&lt;p&gt;Delete the routed-count floor entirely — do not re-pin it. Re-pinning, even to a
low and stable value, keeps the concept, the vocabulary and this ADR alive as
load-bearing, so the next person to move the count reads &lt;em&gt;“DIRECTIVE_043 collapse
tripwire”&lt;/em&gt; and treats it as protective. It was re-pinned 23 times in 63 days,
caught zero regressions, and once failed a build for reading &lt;code dir=&quot;auto&quot;&gt;meta.json&lt;/code&gt; through
the canonical reader the gate exists to reward (KISS audit, R. Douglass,
2026-09-14).&lt;/p&gt;
&lt;p&gt;Remove, in one commit: the &lt;code dir=&quot;auto&quot;&gt;ROUTED_LOAD_META_FLOOR&lt;/code&gt; constant,
&lt;code dir=&quot;auto&quot;&gt;ROUTED_LOAD_META_FLOOR_MARGIN&lt;/code&gt; and its margin assertion, the strict
&lt;code dir=&quot;auto&quot;&gt;assert len(routed) &gt; ROUTED_LOAD_META_FLOOR&lt;/code&gt; anti-vacuity clause,
&lt;code dir=&quot;auto&quot;&gt;EXCLUDED_REL_PATHS&lt;/code&gt;, the &lt;code dir=&quot;auto&quot;&gt;scan_routed_load_meta_calls&lt;/code&gt; scanner and the
&lt;code dir=&quot;auto&quot;&gt;ROUTED_CALLEES&lt;/code&gt; set that fed only it, and all four routed-floor tests (the
predicate, its non-vacuity canary, the real-tree floor, and the mass-allow-list
self-test).&lt;/p&gt;
&lt;p&gt;The &lt;strong&gt;real invariant stays a hard gate, unchanged&lt;/strong&gt;: the inline-read ceiling
(&lt;code dir=&quot;auto&quot;&gt;INLINE_META_READ_FLOOR = 7&lt;/code&gt;), its margin (mechanic 2 — the actual non-vacuity
guard: the live count must stay within &lt;code dir=&quot;auto&quot;&gt;MARGIN&lt;/code&gt; of the ceiling, so a broken or
empty scan reds), the allowlist’s composite-key &lt;code dir=&quot;auto&quot;&gt;{key, rationale, issue}&lt;/code&gt;
requirement, and its stale-entry detection (&lt;code dir=&quot;auto&quot;&gt;allowlist_keys - live_keys&lt;/code&gt; non-empty
fails). That trio &lt;em&gt;is&lt;/em&gt; #4315’s option 3 — “no inline &lt;code dir=&quot;auto&quot;&gt;json.load&lt;/code&gt; of &lt;code dir=&quot;auto&quot;&gt;meta.json&lt;/code&gt;
outside the reader family” — and it already exists, already passes, and has cost
nothing to maintain. The FR-010 single-decoder gate and the independent
&lt;code dir=&quot;auto&quot;&gt;_ACCOUNTED_SITES&lt;/code&gt; census (&lt;code dir=&quot;auto&quot;&gt;test_meta_fail_closed_full_census_contract.py&lt;/code&gt;) retain
the de-routing cover the routed floor claimed.&lt;/p&gt;
&lt;p&gt;Doctrinal note: the gate remains non-vacuous under DIRECTIVE_043 after deletion —
its non-vacuity was always carried by the ceiling’s margin (mechanic 2), never by
the routed floor. Nothing that prevents vacuous passage is removed.&lt;/p&gt;
&lt;h3 id=&quot;2-test_golden_count_ban--retire-and-sweep-the-annotations&quot;&gt;2. &lt;code dir=&quot;auto&quot;&gt;test_golden_count_ban&lt;/code&gt; → &lt;strong&gt;retire, and sweep the annotations&lt;/strong&gt;&lt;/h3&gt;
&lt;p&gt;Delete &lt;code dir=&quot;auto&quot;&gt;tests/architectural/test_golden_count_ban.py&lt;/code&gt; &lt;strong&gt;in full&lt;/strong&gt; — the ceiling
guard, the classifier, the scanner, their unit tests and the &lt;code dir=&quot;auto&quot;&gt;--emit-inventory&lt;/code&gt;
entrypoint — together with &lt;code dir=&quot;auto&quot;&gt;_golden_count_baseline.json&lt;/code&gt;, the module’s
&lt;code dir=&quot;auto&quot;&gt;shape_guard_membership.yaml&lt;/code&gt; row, and
&lt;code dir=&quot;auto&quot;&gt;test_shape_guard_demotion_does_not_raise_under_a_manufactured_breach&lt;/code&gt; (which
asserts the &lt;code dir=&quot;auto&quot;&gt;shape-guard&lt;/code&gt; class is non-empty and so cannot outlive the row).&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Operator decision, 2026-09-14:&lt;/strong&gt; an earlier draft of this verdict retained the
classifier behind &lt;code dir=&quot;auto&quot;&gt;--emit-inventory&lt;/code&gt;. The operator chose wholesale deletion — if
the annotations are swept, the vocabulary that produced them goes too; the
inventory is reconstructible from git history and nothing in-tree consumes it.
This ADR records the reversal rather than letting the decision record and the code
disagree.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Explicitly NOT deleted&lt;/strong&gt;, against the earlier draft:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;code dir=&quot;auto&quot;&gt;shape_guard_membership.yaml&lt;/code&gt; and the rest of &lt;code dir=&quot;auto&quot;&gt;test_shape_guard_membership.py&lt;/code&gt; —
see the Correction above. Only the golden-count row and the one demotion-specific
test go. The &lt;code dir=&quot;auto&quot;&gt;shape-guard&lt;/code&gt; class is left defined but empty.&lt;/li&gt;
&lt;li&gt;&lt;code dir=&quot;auto&quot;&gt;kitty-specs/test-suite-friction-remediation-01KXDKBX/golden-count-inventory.md&lt;/code&gt; —
it lives in a &lt;strong&gt;prior mission’s committed dossier&lt;/strong&gt;, which the charter treats as
an immutable snapshot. Retiring live machinery and rewriting a landed mission’s
artifacts are different acts.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;One fold wholesale deletion forces that retaining the scanner would not:
&lt;code dir=&quot;auto&quot;&gt;test_gate_remedy_presence.py:188&lt;/code&gt; registers
&lt;code dir=&quot;auto&quot;&gt;tests/architectural/test_golden_count_ban.py::ratchet_violations&lt;/code&gt; as a
&lt;strong&gt;known-good remedy exemplar&lt;/strong&gt; — deliberately chosen as one “this WP did not
author” — and reads the file with &lt;code dir=&quot;auto&quot;&gt;read_text()&lt;/code&gt; at &lt;code dir=&quot;auto&quot;&gt;:200&lt;/code&gt;, so deletion raises
&lt;code dir=&quot;auto&quot;&gt;FileNotFoundError&lt;/code&gt; rather than failing an assertion. Repoint that entry at another
unauthored content-anchored remedy; do not simply drop it, or the property check
loses the independence the entry exists to provide.&lt;/p&gt;
&lt;p&gt;Sweep the &lt;strong&gt;387 &lt;code dir=&quot;auto&quot;&gt;# golden-count: cardinality-is-contract&lt;/code&gt; annotations across 194
test files&lt;/strong&gt; in the same mission. Leaving them would preserve a toll contributors
demonstrably keep paying after the charge is gone — three PRs did so in the seven
days after the demotion alone.&lt;/p&gt;
&lt;p&gt;This supersedes &lt;a href=&quot;https://github.com/spec-kitty/spec-kitty/issues/3977&quot;&gt;#3977&lt;/a&gt; and
&lt;a href=&quot;https://github.com/spec-kitty/spec-kitty/issues/3927&quot;&gt;#3927&lt;/a&gt;, both of which
report the pre-demotion hard-red that no longer occurs. #3977’s own disposition —
&lt;em&gt;“Do not re-freeze the baseline”&lt;/em&gt; — is honoured: the baseline is removed, not
re-frozen.&lt;/p&gt;
&lt;h3 id=&quot;3-marker-baseline--already-retired-no-action&quot;&gt;3. Marker baseline → &lt;strong&gt;already retired; no action&lt;/strong&gt;&lt;/h3&gt;
&lt;p&gt;Recorded so #4315 can close its scope item rather than leaving it open against a
file that has not existed since 2026-08-12.&lt;/p&gt;
&lt;h3 id=&quot;4-shard-maps--keep-the-partition-properties-retire-the-archnext-marker-family&quot;&gt;4. Shard maps → &lt;strong&gt;keep the partition properties; retire the &lt;code dir=&quot;auto&quot;&gt;arch&lt;/code&gt;/&lt;code dir=&quot;auto&quot;&gt;next&lt;/code&gt; marker family&lt;/strong&gt;&lt;/h3&gt;
&lt;p&gt;The completeness &lt;em&gt;properties&lt;/em&gt; are the right instrument and stay. What is retired
is the substrate they police, because nothing selects it:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;code dir=&quot;auto&quot;&gt;tests/_arch_shard_map.py&lt;/code&gt;, &lt;code dir=&quot;auto&quot;&gt;tests/_next_shard_map.py&lt;/code&gt;, &lt;code dir=&quot;auto&quot;&gt;tests/_shard_registry.py&lt;/code&gt;;&lt;/li&gt;
&lt;li&gt;&lt;code dir=&quot;auto&quot;&gt;tests/test_shard_registry_fallback.py&lt;/code&gt;;&lt;/li&gt;
&lt;li&gt;&lt;code dir=&quot;auto&quot;&gt;tests/architectural/test_arch_shard_marker_completeness.py&lt;/code&gt;;&lt;/li&gt;
&lt;li&gt;the six &lt;code dir=&quot;auto&quot;&gt;arch_shard_*&lt;/code&gt; / &lt;code dir=&quot;auto&quot;&gt;next_shard_*&lt;/code&gt; registrations in &lt;code dir=&quot;auto&quot;&gt;pytest.ini&lt;/code&gt;;&lt;/li&gt;
&lt;li&gt;the registration + marker-application block in &lt;code dir=&quot;auto&quot;&gt;tests/conftest.py&lt;/code&gt;;&lt;/li&gt;
&lt;li&gt;&lt;code dir=&quot;auto&quot;&gt;test_exemption_registry_ratchet.py::test_new_arch_test_files_are_shard_registered&lt;/code&gt;,
a T057/C-006 check of a property that ceases to exist.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&lt;strong&gt;Explicitly retained&lt;/strong&gt;, as different mechanisms with live consumers:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;code dir=&quot;auto&quot;&gt;tests/architectural/test_fast_tier_marker_completeness.py&lt;/code&gt; — tier markers,
derived live from the &lt;code dir=&quot;auto&quot;&gt;Makefile&lt;/code&gt;, genuinely select &lt;code dir=&quot;auto&quot;&gt;make test-fast&lt;/code&gt;;&lt;/li&gt;
&lt;li&gt;&lt;code dir=&quot;auto&quot;&gt;tests/architectural/test_module_shard_registry.py&lt;/code&gt; and
&lt;code dir=&quot;auto&quot;&gt;.github/ci-module-registry.yml&lt;/code&gt; — the live directory-based CI sharding, whose
own non-vacuity floor is &lt;em&gt;derived&lt;/em&gt; (&lt;code dir=&quot;auto&quot;&gt;ceiling // 2&lt;/code&gt;), not pinned.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&lt;strong&gt;Folds the deletions force — corrected count and corrected mechanics.&lt;/strong&gt; An earlier
draft named “two folds”, one of them wrongly. The real shape, measured in
&lt;code dir=&quot;auto&quot;&gt;pyproject.toml&lt;/code&gt;:&lt;/p&gt;
&lt;p&gt;&lt;code dir=&quot;auto&quot;&gt;[tool.ruff.format].exclude&lt;/code&gt; carries an entry for &lt;strong&gt;four&lt;/strong&gt; of the files these two
verdicts delete — &lt;code dir=&quot;auto&quot;&gt;tests/architectural/test_golden_count_ban.py&lt;/code&gt; (&lt;code dir=&quot;auto&quot;&gt;:1087&lt;/code&gt;),
&lt;code dir=&quot;auto&quot;&gt;tests/_shard_registry.py&lt;/code&gt; (&lt;code dir=&quot;auto&quot;&gt;:974&lt;/code&gt;),
&lt;code dir=&quot;auto&quot;&gt;tests/test_shard_registry_fallback.py&lt;/code&gt; (&lt;code dir=&quot;auto&quot;&gt;:2757&lt;/code&gt;) and
&lt;code dir=&quot;auto&quot;&gt;tests/architectural/test_arch_shard_marker_completeness.py&lt;/code&gt; (&lt;code dir=&quot;auto&quot;&gt;:1060&lt;/code&gt;). All four
entries must be removed, or &lt;code dir=&quot;auto&quot;&gt;test_ruff_format_exclude_ratchet.py::test_every_exclude_entry_exists_on_disk&lt;/code&gt;
reds. The earlier draft named only the last of the four.&lt;/p&gt;
&lt;p&gt;The pinned-count coupling was also stated backwards. &lt;code dir=&quot;auto&quot;&gt;_BASELINE_EXCLUDE_COUNT&lt;/code&gt;
(2809) is enforced as &lt;code dir=&quot;auto&quot;&gt;len(entries) &amp;#x3C;= baseline&lt;/code&gt;, so &lt;strong&gt;removal never requires
touching it&lt;/strong&gt; — only growth does. The mandatory half is the entry removal; the
count edit is optional. Stating it the other way round would have sent an
implementer looking for a coupling that does not exist in the direction this work
moves, while leaving the one that does bite unnamed for three of four entries.&lt;/p&gt;
&lt;p&gt;&lt;code dir=&quot;auto&quot;&gt;tests/_arch_shard_map.py&lt;/code&gt; and &lt;code dir=&quot;auto&quot;&gt;tests/_next_shard_map.py&lt;/code&gt; carry no exclude entry.
The two duplicate &lt;code dir=&quot;auto&quot;&gt;next&lt;/code&gt; rows and the dead &lt;code dir=&quot;auto&quot;&gt;_gate_read_callshape.py&lt;/code&gt; row disappear
with the tables.&lt;/p&gt;
&lt;h2 id=&quot;consequences&quot;&gt;Consequences&lt;/h2&gt;
&lt;p&gt;&lt;strong&gt;Good&lt;/strong&gt;&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;Removes a hard gate sitting at &lt;strong&gt;zero headroom&lt;/strong&gt; whose failure mode is punishing
correct behaviour, and that has never caught a regression in 23 re-pins.&lt;/li&gt;
&lt;li&gt;Removes ~2,500 lines of gate machinery and ~390 annotation sites with a measured
lifetime catch record of zero.&lt;/li&gt;
&lt;li&gt;Eliminates an estimated ~40 of the last ~70 baseline-maintenance commits across
these gates.&lt;/li&gt;
&lt;li&gt;Restores single-canonical-authority: &lt;code dir=&quot;auto&quot;&gt;tests/architectural/_baselines.yaml&lt;/code&gt;
becomes the only frozen-baseline authority.&lt;/li&gt;
&lt;li&gt;Closes #3977 and #3927 as superseded; resolves #4315’s four scope items.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&lt;strong&gt;Bad / accepted risk&lt;/strong&gt;&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;The routed census may now drift far from its floor unobserved. Accepted: the
&lt;em&gt;defect&lt;/em&gt; — inline reads regrowing outside the reader family — is caught by the
retained ceiling + allowlist, which is the gate with the real catch record. A
falling routed count with a flat inline ceiling is a refactor, not a regression.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Deleting the routed floor removes the routing-evidence signal entirely —
recorded, not papered over (adversarial squad, 2026-09-14).&lt;/strong&gt; The pre-deletion
gate had &lt;code dir=&quot;auto&quot;&gt;EXCLUDED_REL_PATHS&lt;/code&gt; remove &lt;code dir=&quot;auto&quot;&gt;src/specify_cli/mission_metadata.py&lt;/code&gt; and
&lt;code dir=&quot;auto&quot;&gt;src/specify_cli/task_utils/support.py&lt;/code&gt; from the &lt;em&gt;inline&lt;/em&gt; scan; those two files
held 17 of the ~157 routed sites, and inside them the routed floor had been the
only cover. With the floor deleted — and &lt;code dir=&quot;auto&quot;&gt;EXCLUDED_REL_PATHS&lt;/code&gt; gone with it,
having measured to hide zero &lt;em&gt;inline&lt;/em&gt; sites — there is no routed-count
enforcement anywhere. Accepted: the de-routing defect (replacing canonical-reader
delegations with hand-rolled parsing) is covered by the independent
&lt;code dir=&quot;auto&quot;&gt;_ACCOUNTED_SITES&lt;/code&gt; exact-equality census in
&lt;code dir=&quot;auto&quot;&gt;test_meta_fail_closed_full_census_contract.py&lt;/code&gt; (keyed on &lt;code dir=&quot;auto&quot;&gt;load_meta_fail_closed&lt;/code&gt;,
floor-independent) plus review; and the routed floor’s own record is 23 re-pins,
zero regression catches, both historical drops answered by lowering the pin
rather than investigating. An earlier draft planned to keep a small pinned
decode-count &lt;em&gt;inside&lt;/em&gt; the two files; outright deletion abandons that mitigation
as more of the vocabulary the audit calls to remove.&lt;/li&gt;
&lt;li&gt;Retiring the golden-count ceiling means a future &lt;code dir=&quot;auto&quot;&gt;len(X) == N&lt;/code&gt; regrowth is caught
by review rather than CI. Accepted: it has always been caught by review, since
the gate caught nothing in its lifetime and has been advisory since 2026-09-07.&lt;/li&gt;
&lt;li&gt;Retiring the shard markers discards the balance-control substrate if
marker-based sharding is ever reinstated. Accepted: the current substrate is
demonstrably unmaintained (two duplicate rows it cannot see), and live sharding
has moved to a directory-based registry. Reinstating would start from that
registry, not from these tables.&lt;/li&gt;
&lt;li&gt;The annotation sweep touches 194 test files. Accepted deliberately, to stop the
cargo-culted toll; it is comment-only and behaviour-preserving.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&lt;strong&gt;Corrections to this ADR’s own claims (post-spec adversarial squad, 2026-09-14)&lt;/strong&gt;&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;“The real invariant stays a hard gate” overstates the retained scanner.&lt;/strong&gt; A
squad lens built a scratch &lt;code dir=&quot;auto&quot;&gt;src/&lt;/code&gt; tree and ran the real &lt;code dir=&quot;auto&quot;&gt;scan_inline_meta_reads&lt;/code&gt;
against eight plausible new inline reads plus one canonical control: &lt;strong&gt;one
caught, eight evaded.&lt;/strong&gt; The evasions include &lt;code dir=&quot;auto&quot;&gt;json.loads(path.read_bytes())&lt;/code&gt;
(the read-base matcher only knows &lt;code dir=&quot;auto&quot;&gt;read_text&lt;/code&gt;/&lt;code dir=&quot;auto&quot;&gt;open&lt;/code&gt;), &lt;code dir=&quot;auto&quot;&gt;Path(dir, &quot;meta.json&quot;)&lt;/code&gt;
and &lt;code dir=&quot;auto&quot;&gt;os.path.join(...)&lt;/code&gt; (the join matcher requires a &lt;code dir=&quot;auto&quot;&gt;/&lt;/code&gt; &lt;code dir=&quot;auto&quot;&gt;BinOp&lt;/code&gt;), a module-level
&lt;code dir=&quot;auto&quot;&gt;META_NAME&lt;/code&gt; constant (it requires a literal), an interprocedural one-line helper,
and &lt;code dir=&quot;auto&quot;&gt;yaml.safe_load(...)&lt;/code&gt; — which parses JSON correctly and is idiomatic in this
codebase. &lt;strong&gt;None of this is a regression introduced here&lt;/strong&gt; — all eight pass today
with the margin intact — but the retained gate closes one AST &lt;em&gt;spelling&lt;/em&gt; of the
defect class, not the class. The verdicts stand; the confidence behind them is
hereby narrowed to what was measured.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;A second, tighter de-routing detector already exists and was unnamed.&lt;/strong&gt;
&lt;code dir=&quot;auto&quot;&gt;tests/specify_cli/test_meta_fail_closed_full_census_contract.py&lt;/code&gt; keeps an
independent &lt;code dir=&quot;auto&quot;&gt;_ACCOUNTED_SITES&lt;/code&gt; ledger checked for &lt;strong&gt;exact equality&lt;/strong&gt; against a
live scan, keyed on &lt;code dir=&quot;auto&quot;&gt;load_meta_fail_closed&lt;/code&gt;. It does not depend on
&lt;code dir=&quot;auto&quot;&gt;ROUTED_LOAD_META_FLOOR&lt;/code&gt; and is unaffected by this ADR — so the de-routing class
retains cover this ADR did not credit.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Sequencing dependency (discharged).&lt;/strong&gt; This ADR landed first (#4323) and the
implementing mission (#4359) rebased onto it, so the mission branch carries this
file and flips it &lt;code dir=&quot;auto&quot;&gt;Proposed&lt;/code&gt; → &lt;code dir=&quot;auto&quot;&gt;Accepted&lt;/code&gt; in a dedicated commit, separate from
its three workstream commits. Accepted on the operator’s direction to land the
flip with the implementation rather than as a follow-up.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&lt;strong&gt;Neutral&lt;/strong&gt;&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;The ~2,500-line reduction is not itself the justification — the catch record is.&lt;/li&gt;
&lt;li&gt;This ADR changes no doctrine. DIRECTIVE_043 and both tactics remain in force and
unedited; verdict 1 deletes a redundant floor whose non-vacuity role the ceiling’s margin already carries, and verdicts 2 and 4
retire gates whose defect classes no longer exist.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id=&quot;follow-through&quot;&gt;Follow-through&lt;/h2&gt;
&lt;p&gt;Implementation runs as a governed mission, sequenced so &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt; stays green at
every intermediate commit: the routed-floor deletion and each retirement are independent
and land separately. If any verdict’s implementation surfaces a consumer this ADR
did not name, that is a finding against this ADR, not a licence to improvise
around it.&lt;/p&gt;
&lt;h2 id=&quot;references&quot;&gt;References&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;Issues: &lt;a href=&quot;https://github.com/spec-kitty/spec-kitty/issues/4315&quot;&gt;#4315&lt;/a&gt; (primary),
&lt;a href=&quot;https://github.com/spec-kitty/spec-kitty/issues/1931&quot;&gt;#1931&lt;/a&gt; (epic),
&lt;a href=&quot;https://github.com/spec-kitty/spec-kitty/issues/2913&quot;&gt;#2913&lt;/a&gt; (sibling +
binding operator ruling), &lt;a href=&quot;https://github.com/spec-kitty/spec-kitty/issues/3458&quot;&gt;#3458&lt;/a&gt;
(demotion precedent), &lt;a href=&quot;https://github.com/spec-kitty/spec-kitty/issues/3977&quot;&gt;#3977&lt;/a&gt;
and &lt;a href=&quot;https://github.com/spec-kitty/spec-kitty/issues/3927&quot;&gt;#3927&lt;/a&gt; (superseded),
&lt;a href=&quot;https://github.com/spec-kitty/spec-kitty/issues/3285&quot;&gt;#3285&lt;/a&gt; (marker-baseline
retirement).&lt;/li&gt;
&lt;li&gt;Doctrine: &lt;code dir=&quot;auto&quot;&gt;packs/built-in/directives/043-close-defect-class-by-construction.directive.yaml&lt;/code&gt;,
&lt;code dir=&quot;auto&quot;&gt;packs/built-in/tactics/architectural-gate-non-vacuity.tactic.yaml&lt;/code&gt;,
&lt;code dir=&quot;auto&quot;&gt;packs/built-in/tactics/frozen-baseline-shrink-only-ratchet.tactic.yaml&lt;/code&gt;,
&lt;code dir=&quot;auto&quot;&gt;packs/built-in/procedures/post-merge-arch-gate-adjudication.procedure.yaml&lt;/code&gt;.&lt;/li&gt;
&lt;/ul&gt;</content>
</entry>
