docs: promote verified H4 L5 closure state

[JoonasHuhta]

Promotes the merged H4 evidence package into the project's canonical research state. Every scientific number is derived from the merged package (docs/evidence/h4-l5-full-closure-2026-08/), not from memory.

MATH_CLAIMS.md
New row 111 (next free id). Route (c) at L=5, full coding family: every
uniform coding g : S6 -> S3^5, all 243^6 = 205,891,132,094,649, is eliminated
by at least one of two preregistered finite tests. Stage A: block-aligned
Parikh obstruction M_g.d = 0 from h6^9(a), m in [2,120], K = 5m in [10,600],
first 98,415 coded symbols. Stage B: abelian square K in [6,40] in the first
3,645 coded symbols of g(h6^6(a)). Stage A retained 5,153,928 profiles =
3,316,540,933,500 codings; Stage B processed and rejected all
3,316,540,933,500, leaving 0 survivors across 2219/2219 shards. Status
COMPUTED (Level 1). Non-vacuity is stated with the exact counts only, not
with hand-derived percentages.

The row carries the evidence package's own explicit non-claims (not Makela's
conjecture, not non-uniform morphisms, not L >= 6, not all ternary
constructions, not other source morphisms), the 98,415-vs-3,645 combined
window note, and all three open provenance caveats: per-shard executable
identity not recorded across resumes; real-survivor validation path never
naturally exercised (synthetic alert self-test only); Stage-A artifact has no
second physical backup.

Row 49 gains a cross-reference sentence only. Row 111 is a separate
full-family closure for L=5, NOT a replacement: row 49 remains the authority
for its own differently scoped computation (S3-canonical classes, its own
source windows, wider K in [6,100]) and for L = 1..4. Row 49's claim text,
status and both dates are unchanged.

KNOWLEDGE_STATE.md one table line in section 3, citing row 111.
CURRENT_FOCUS.md H4 block replaced: COMPLETE, promoted, no longer active
work; next work is B22 exactly as preregistered. Kill-depth
is explicitly outside B22, and "shallow L=5 kills imply L=6
is dead" is recorded as an exploratory hypothesis only.
NEXT_STEP.md top staleness/subordination banner only, 19 lines added and
0 removed; historical body byte-identical.

Two derived-data couplings the repository enforces, both mechanical:

RESEARCH_CONTEXT.md 114 rows -> 115 rows. Required by
scripts/check-claims-drift.js. No H4 prose added.
explorer.html claims-data block resynced by the canonical exporter
(node src/claims-export.js). rowCount 114 -> 115 and
COMPUTED 77 -> 78 are exporter-generated, never typed. The
diff is 1 line changed; everything outside the generated
<script id="claims-data"> block hashes identically to base
(71fb712c...). No visible copy, UI, CSS, behaviour or route
change. QUOTABLE_FACTS is untouched, so row 111 is not
quotable and no public figure changed.

Deliberately NOT changed: the H4 evidence package, the B22 preregistration, EPISTEMIC_DISCIPLINE.md, the pre-existing duplicate ledger row id 84, and the CRLF/manifest portability issue. The last two are recorded as separate deferred items for later read-only audit.

Verification: node tests/test.js (42/0), node scripts/check-claims-drift.js (18/0), node src/claims-export.js --check (exit 0) - raw output inspected for all three.