docs: align prose with UK English

docs/governance/CHARTER.md

@@ -58,7 +58,7 @@ Eval:
 - `retain` means keep the family in the active lane and keep gathering baseline
   evidence under the current rule
 - `evict` means make the later upstream runtime correction because the failure
-  family has stabilized enough to earn removal
+  family has stabilised enough to earn removal
 - do not keep re-judging a known bad family forever once it has earned
   eviction
 - one eval focus at a time

docs/governance/DECISIONS.md

@@ -132,9 +132,9 @@ If a decision crosses layers, say so plainly instead of flattening the method in
   - use one shared style-signal resource across all prompt types
   - let prompt type control the reasoning lane, not the flavour pool
   - treat the style signals as compositional cues rather than as a fixed lexical pool
-  - keep the model free to synthesize beyond the literal words provided
+  - keep the model free to synthesise beyond the literal words provided
   - this shape was explicitly clarified through human direction during the tone pass
-- Why: The prompt type should control the reasoning lane, while the model remains free to synthesize beyond the literal words provided.
+- Why: The prompt type should control the reasoning lane, while the model remains free to synthesise beyond the literal words provided.
 
 ## D-009: Hold the baseline and learn from failures
 
@@ -363,7 +363,7 @@ If a decision crosses layers, say so plainly instead of flattening the method in
     - immediate judgment
     - next product
   - but treat a longer serial chunk as the minimum useful surface for a live checkpoint
-  - use `25+` rows as the default minimum useful chunk before summarizing progress
+  - use `25+` rows as the default minimum useful chunk before summarising progress
   - treat `50-100` rows, or about one hour of serial lane work, as the real long-run surface
   - keep extra `when` pressure in the mix because that lane is the most useful stress test for the current coherence rule
 - Why: The value is in the data. Tiny taste-check runs make the method look cleaner than it really is and do not produce enough pressure to learn from the current coherence threshold.
@@ -551,7 +551,7 @@ If a decision crosses layers, say so plainly instead of flattening the method in
   - `retain` means the failure family stays in the active lane while the fixed
     rule keeps taking pressure
   - reserve `evict` for the later upstream runtime or boundary correction made
-    because that evidence has stabilized into one known bad family
+    because that evidence has stabilised into one known bad family
   - do not keep re-judging a family forever once it has clearly moved from
     active evidence into known residue
   - do not claim a `when` eviction fix yet while the current Beta `4.1` lane is
@@ -735,7 +735,7 @@ If a decision crosses layers, say so plainly instead of flattening the method in
   - `git diff --check`
 - Why: Probaboracle's startup ritual should end with the same explicit operator
   pause as the rest of the toy family. The docs list matters most at the final
-  stop sign, where the next instance has to read, summarize, and name exactly
+  stop sign, where the next instance has to read, summarise, and name exactly
   one active kernel before doing repo work.
 
 ## D-046: Align the full start/end operator surface with the repo family
@@ -829,7 +829,7 @@ If a decision crosses layers, say so plainly instead of flattening the method in
     - tone contract
     - output constraints
     - lane guardrails
-  - remove content-led scaffolds that oversteer the line toward repeated wording
+  - remove content-led scaffolds that oversteer the line towards repeated wording
     or stock closures
   - do not reintroduce a disguised word bank through examples, favourite
     closers, or preferred stock phrases
@@ -843,7 +843,7 @@ If a decision crosses layers, say so plainly instead of flattening the method in
     - `D-010`
     - `docs/governance/SESSION_HANDOFF.md`
   - confirm the live instruction path no longer hard-codes phrase examples as
-    favored output content
+    favoured output content
 - Why: The durable repo rules already say that style signals are cues, not a
   word bank, and that content-led cues invite repetition and drift. The live
   runtime surface had not fully caught up: it still carried explicit phrase
@@ -959,7 +959,7 @@ If a decision crosses layers, say so plainly instead of flattening the method in
   local operator surface hid some checks behind npm commands and carried a
   stale duplicate end routine. The repo family needs a compact, explicit
   tooling surface that can be mirrored by the other toys without widening
-  runtime or eval behavior.
+  runtime or eval behaviour.
 
 ## D-053: Beta 6.0 judges one fixed-prompt pulse
 

docs/governance/SESSION_HANDOFF.md

@@ -1,6 +1,6 @@
 # Session Handoff
 
-Last updated: 2026-05-22
+Last updated: 2026-05-24
 
 ## Start Here
 

docs/research/BETA_1_PRODUCT_FIT.md

@@ -34,7 +34,7 @@ flowchart LR
 
 ## What It Showed
 
-- pushing the tone toward deadpan, vague, and unhelpful
+- pushing the tone towards deadpan, vague, and unhelpful
 - finding obvious misses quickly
 - shaping the product voice
 

docs/research/BETA_5_RETAIN_OR_EVICT.md

@@ -98,7 +98,7 @@ The post-evict confirmation rerun then covered rows `4098-4197`:
 - `100` rows total
 - `97 pass / 3 fail / 0 pending`
 
-The note mix shifted hard toward clean passes:
+The note mix shifted hard towards clean passes:
 
 - `83` `good scheduling refusal`
 - `14` `plain when line`

docs/runtime/ARCHITECTURE.md

@@ -110,7 +110,7 @@ The public generation and eval-shape diagrams live in `docs/diagrams/PIPELINE.md
   - `pass`
   - `fail`
 - Row-level baselines may decide `retain / evict` after a failure family
-  stabilizes.
+  stabilises.
 - Beta `6.0` rows are not product-judged; they feed one fixed-prompt pulse
   verdict.
 - Runtime directions describe the target reasoning shape rather than accumulating long restriction lists.