fix(framework): key §5 lookup on rendered bullets, not the token (mos-dt round-2)

§5 sent the agent to read direct|friendly|formal in USER.md, but the builder
renders prose bullets, not the token — the documented lookup could not key on
the shipped file. Table now keys on the leading bullet USER.md actually
contains. Also: 'concise, technical' -> 'concise, structured' (drop the round-1
residual value name from a rule-9 guide). Docs-only, no code, no scope growth.

Written-by: jarvis (dragon-lin)
Co-Authored-By: Claude Fable 5 <[email protected]>

packages/mosaic/framework/defaults/AGENTS.md

@@ -39,6 +39,7 @@ overwritten on upgrade. (Layer model: `constitution/LAYER-MODEL.md`.)
 | TypeScript strict typing                           | `guides/TYPESCRIPT.md`             |
 | QA / test strategy                                 | `guides/QA-TESTING.md`             |
 | Documentation (any code/API/auth/infra change)     | `guides/DOCUMENTATION.md`          |
+| Writing style (docs, comms, any prose)             | `guides/WRITING-STYLE.md`          |
 | Secrets / vault usage                              | `guides/VAULT-SECRETS.md`          |
 | Tool/credential reference (service CLIs, wrappers) | `guides/TOOLS-REFERENCE.md`        |
 | Memory protocol (OpenBrain capture/recall)         | `guides/MEMORY.md`                 |

packages/mosaic/framework/defaults/STANDARDS.md

@@ -27,6 +27,14 @@ Master/slave model:
 - Do not perform destructive git/file actions without explicit instruction.
 - Browser automation (Playwright, Cypress, Puppeteer) MUST run in headless mode. Never launch a visible browser — it collides with the user's display and active session.
 
+### Output standards (writing + code)
+
+- Technical documentation follows **MOS-STE** (Mosaic Simplified Technical English — an adapted ASD-STE100 profile): short sentences, one instruction per sentence, active voice, one word per meaning, one term per concept. Full rules: `~/.config/mosaic/guides/WRITING-STYLE.md`.
+- Apply MOS-STE **hardest to verification artifacts** (acceptance criteria, witness predicates, gate/alarm conditions). There an ambiguous term produces a false green, not just a confused reader.
+- Source code follows the **Google Style Guide** for the language.
+- User-facing comms follow the user's declared `communicationStyle` in `USER.md` "Communication Preferences" (`direct` | `friendly` | `formal`, default `direct`); `guides/WRITING-STYLE.md` §5 maps each value to output. The documentation standard does not change with user preference.
+- **Carve-out:** MOS-STE does NOT apply to content that must carry a specific human voice (letters, personal or marketing prose, voice-matched output). A declared voice profile wins.
+
 ### Secrets handling (HARD RULE)
 
 - Vault is the canonical source-of-truth for every secret in every environment. No exceptions.

packages/mosaic/framework/guides/WRITING-STYLE.md

@@ -0,0 +1,134 @@
+# Writing Style Standard — MOS-STE (MANDATORY)
+
+This guide defines how agents write. It sets one style standard per output type.
+It is written in the standard it defines, as a worked example.
+
+**Adapted, not compliant.** MOS-STE (Mosaic Simplified Technical English) is an
+adapted profile of ASD-STE100. Mosaic does not license or certify against
+ASD-STE100. Mosaic uses the load-bearing rules and fits them to agent work. This
+is the same stance Mosaic takes toward DO-178B/C: use the rigor, do not claim the
+certification.
+
+## Scope — which standard governs which output
+
+| Output type | Standard |
+|---|---|
+| Technical documentation (READMEs, runbooks, PRDs, procedures, ADRs, guides, acceptance criteria, design docs) | **MOS-STE** (this guide) |
+| Source code and code comments | **Google Style Guide** for the language (§4) |
+| Inter-agent comms | MOS-STE by default (concise, structured) |
+| User-facing comms | **Per-user style choice** — read `USER.md` "Communication Preferences" (§5) |
+| End-user prose the user owns (marketing, letters, personal writing, voice-matched content) | The user's declared voice. MOS-STE does NOT apply. |
+
+**The user-voice carve-out is absolute.** Do not apply MOS-STE to content that
+must carry a specific human voice (for example a cover letter, a personal
+message, or marketing copy). That content needs the user's voice. MOS-STE would
+damage it. When a project declares a voice profile, that profile wins.
+
+## 1. Why one standard
+
+Agent documentation drifts across projects. Different agents use different terms,
+sentence styles, and structures for the same concept. Readers lose time.
+Assumptions hide in ambiguous prose. One standard gives agents a clear target. It
+gives reviewers a clear test.
+
+## 2. Where MOS-STE matters most — verification artifacts
+
+Apply MOS-STE hardest to acceptance criteria, witness predicates, gate
+definitions, and alarm conditions. In prose, an ambiguous term produces a
+confused reader. In a verification artifact, an ambiguous term produces a false
+green — a check that passes without testing the claim.
+
+The one-term-one-concept rule (rule 9) is the guard. When one word names two
+concepts in one predicate, the check can test the wrong concept and still pass.
+
+**Worked failure.** A rename used a witness predicate with three clauses: ref A
+present, ref B absent, tip committed from this host. Every clause tested the git
+*ref* (the channel). The claim under test was about a *field inside the payload*.
+The word "beacon" named two concepts in one sentence. Deleting ref B was the next
+scheduled step. That step flips the last clause green and certifies a state in
+which the payload still names the wrong host. The predicate was one planned action
+away from a false green on its normal path. The payload field was never tested.
+
+Rule: when N failure modes share one observable, the observable is not a
+diagnostic. In a verification artifact, that ambiguity does not confuse a reader —
+it certifies the defect.
+
+## 3. MOS-STE rules
+
+### 3.1 Sentence rules
+
+1. Keep sentences short. Use 20 words or fewer for a procedure. Use 25 words or
+   fewer for a description. (Reasoning and doctrine prose relaxes this limit —
+   see §3.4. A future lint enforces §3.1, not §3.4.)
+2. Write one instruction per sentence. In a procedure, give one command per step.
+3. Use the active voice. Write "Run the script." Do not write "The script should
+   be run."
+4. Use the imperative for instructions. Start the sentence with the verb.
+5. Use simple verb tenses. Prefer the present tense. Avoid the perfect and
+   progressive tenses when a simple tense works.
+6. Do not use an `-ing` form when it makes the meaning unclear.
+7. Write positive statements. State what to do, not only what to avoid.
+
+### 3.2 Word rules
+
+8. Use one word for one meaning. Do not use the same word in two senses.
+9. Use one term for one concept. Do not use synonyms for variety. Example: choose
+   `secret`, `credential`, or `key` for each concept, and keep it.
+10. Use articles (`a`, `the`). Do not drop words to save space.
+11. Keep an approved-terms glossary per project. Add each domain noun and each
+    chosen verb. Technical names (for example `Vault`, `cgroup`, `systemd`) are
+    always allowed.
+12. Define an abbreviation at its first use. Then use it consistently.
+
+### 3.3 Structure rules
+
+13. Use a list for parallel items or sequential steps. Do not put them in one long
+    sentence.
+14. Use a table for data with more than two dimensions.
+15. Use parallel structure in headings and steps.
+16. Repeat the noun. Do not use a pronoun when the reference is unclear.
+
+### 3.4 Adaptation notes (where MOS-STE deviates from ASD-STE100, and why)
+
+- **No licensed dictionary.** ASD-STE100 ships a controlled dictionary under
+  copyright. MOS-STE uses per-project glossaries instead (rule 11).
+- **Domain terms are allowed.** MOS-STE keeps every term the work needs.
+- **Reasoning prose gets structure, not amputation.** Apply the sentence and word
+  rules to design and doctrine writing. Allow the length a subtle argument needs.
+  Readable-first beats rule-strict when the two conflict.
+
+## 4. Code — Google Style Guide
+
+Write source code to the Google Style Guide for the language (Python, TypeScript,
+Shell, Go, and so on). Match the existing file when a local convention already
+exists. Keep code comments to the MOS-STE sentence and word rules.
+
+## 5. User-facing comms — a per-user choice
+
+Mosaic is multi-user. Different users want different comms styles. The framework
+already carries the selectable setting: `communicationStyle` (`direct` |
+`friendly` | `formal`, default `direct`). `mosaic init` writes it, and the
+builder renders it into the generated `USER.md` "Communication Preferences"
+section. This guide adds the OUTPUT meaning of each value; do not invent new
+values.
+
+The builder renders the style as prose bullets, not the token name, so match on
+the leading bullet the generated `USER.md` actually contains:
+
+| `USER.md` leading bullet | Style | User-facing output |
+|---|---|---|
+| "Direct and concise" | `direct` (default) | MOS-STE structure — short, active, defined terms, tables for overview. |
+| "Warm and conversational" | `friendly` | Warmer register. Full sentences, explain reasoning, fewer tables. |
+| "Professional and structured" | `formal` | Professional and structured. Thorough, with explicit recommendations. |
+
+This setting governs **user-facing comms only**. It does not change the
+documentation standard (§3), which is always MOS-STE regardless of the value.
+
+## 6. Enforcement
+
+- **Now:** human review only. **No mechanical prose check exists today.** The
+  pre-push gate runs typecheck, lint, build, and tests; it inspects no prose.
+  Reviewers check output against the scope table and the MOS-STE rules by hand.
+- **Future:** an MOS-STE lint check (built from the §3.1 sentence rules) and a
+  Google-style linter in the pre-push gate. A future linter enforces §3.1, not
+  §3.4 — see the note at rule 1.

packages/mosaic/framework/tools/wake/manifest.txt

@@ -347,23 +347,6 @@
 #              "one targeted call, never a search" (NOT "pins the observed
 #              state") — sequenced AFTER the reseed so the edit itself is a
 #              live delivery test of the fixed gate.
-#   0.6.16 #952 quarantine-audit clean-sweep message named only ONE of the
-#              two unprovable residual classes ("rows without surviving
-#              dead-letter evidence"), so an operator reading the OK concluded
-#              NO evidence exists when evidence can exist and be UNUSABLE: a
-#              surviving dead-letter row whose locator extracts an empty
-#              observed_hash (live specimen: mos-dt seq 13,
-#              bench/malformed-locator-test, "deliberately non-conformant")
-#              can never satisfy the four-field conviction match, because
-#              _record_last_consumed only writes rows with a NON-empty hash.
-#              WORDING-ONLY fix (measurement defect, #951 review finding 1):
-#              the clean-sweep message + the PROVABILITY BOUND comment now
-#              name both classes; the conviction predicate is UNCHANGED.
-#              Test T17 drives the real writer flow and plants the verbatim
-#              live specimen (nested .locators.*, NO observed_hash key) —
-#              never a hand-built flat dead-letter row, which would make the
-#              audit's correct non-conviction look exactly like the defect
-#              under hunt (the #951 review's false-defect near-miss).
 #   0.6.15 #946 the digest's embedded ack watermark covered quarantined
 #              entries: quarantine is a render-time filter (0.6.14), so the
 #              suggested `ack.sh consumed --upto <observed_seq>` stepped the
@@ -392,7 +375,7 @@
 #              (+ test-wake-store-ack.sh T13-T16,
 #              test-wake-digest-quarantine.sh Q12-Q16).
 component=wake
-version=0.6.16
+version=0.6.15
 
 # Watch-list schema this component consumes, and the INCLUSIVE range of
 # schema_version values it supports. A wake-watch-list.json whose schema_version

packages/mosaic/framework/tools/wake/store.sh

@@ -110,13 +110,9 @@ Commands:
                                          found); --repair removes exactly those
                                          rows (atomic, loud). The dead-letter
                                          ledger itself is NEVER modified (it is
-                                         history). TWO residual classes are
-                                         unprovable and never touched (#952):
-                                         rows whose dead-letter evidence was
-                                         pruned away, AND rows whose surviving
-                                         evidence extracts an empty
-                                         observed_hash (it can never satisfy
-                                         the four-field conviction match).
+                                         history). Rows whose dead-letter evidence
+                                         was pruned are NOT provable and are
+                                         never touched.
   cursors                                Print observed_seq / consumed_seq / depth.
 
 Environment:
@@ -567,17 +563,10 @@ cmd_quarantine_sync() {
 # it on (kind, id, observed_seq, observed_hash) AND row.observed_seq <=
 # consumed_seq: the per-key max_by merge means the surviving row's provenance IS
 # that quarantined entry (a healed row differs in seq/hash and never matches).
-# PROVABILITY BOUND — TWO residual classes, both unprovable (#952): (1) a row
-# whose dead-letter evidence was pruned/rotated away — no evidence to convict
-# on; (2) a row whose dead-letter evidence SURVIVES but extracts an empty
-# observed_hash (e.g. a deliberately non-conformant locator: nested
-# .locators.* with no observed_hash key) — evidence exists but can never
-# satisfy the four-field match, because _record_last_consumed only ever writes
-# rows with a NON-empty hash. Neither class is touched — this audit only ever
-# removes what the ledger can convict, and the clean-sweep message names BOTH
-# classes: "no evidence" and "evidence unusable" are different operator
-# conclusions. The dead-letter ledger itself is history and is NEVER modified
-# here.
+# PROVABILITY BOUND: a row whose dead-letter evidence was pruned/rotated away is
+# NOT provable and is never touched — this audit only ever removes what the
+# ledger can convict. The dead-letter ledger itself is history and is NEVER
+# modified here.
 cmd_quarantine_audit() {
   local repair=0
   while [ $# -gt 0 ]; do
@@ -616,7 +605,7 @@ cmd_quarantine_audit() {
       )) | length) > 0)
   ' "$rec" 2>/dev/null || true)"
   if [ -z "$false_rows" ]; then
-    echo "store.sh quarantine-audit: OK — no provably-false consumed-hash rows. Two residual classes are unprovable and were NOT judged: rows whose dead-letter evidence was pruned/rotated away (no evidence to convict on), and rows whose surviving dead-letter evidence extracts an empty observed_hash (evidence exists but can never satisfy the four-field conviction match)."
+    echo "store.sh quarantine-audit: OK — no provably-false consumed-hash rows (rows without surviving dead-letter evidence are not provable and were not judged)"
     return 0
   fi
   local n row

packages/mosaic/framework/tools/wake/test-wake-store-ack.sh

@@ -28,15 +28,6 @@
 #   T16 #946 quarantine-audit: consumed-hashes rows provably false against the
 #      dead-letter ledger are reported (exit 1) and removed only under --repair;
 #      healed rows and the ledger itself are untouched
-#   T17 #952 quarantine-audit clean sweep names BOTH unprovable residual
-#      classes: evidence pruned (no evidence) AND evidence surviving with an
-#      empty observed_hash (evidence unusable). Fixture is the VERBATIM live
-#      specimen (mos-dt lane dead-letter seq 13, bench/malformed-locator-test)
-#      — a nested .locators.* row with NO observed_hash key. Do NOT hand-build
-#      a FLAT dead-letter fixture here: consumed-hashes rows are genuinely
-#      flat, dead-letter rows are genuinely nested, and a flat fixture makes
-#      the audit's CORRECT non-conviction look exactly like the defect under
-#      hunt (this false-defect near-miss actually happened in the #951 review)
 #
 # Isolated: every test runs against a fresh WAKE_STATE_HOME temp dir.
 set -uo pipefail
@@ -689,50 +680,6 @@ echo "== T16: #946 — quarantine-audit: a consumed-hash row matching a dead-let
   grep -qi 'OK' "$TMP_ROOT/t16.ok" || fail_msg "T16: a clean audit must say OK [$(cat "$TMP_ROOT/t16.ok")]"
 ) && ok
 
-echo "== T17: #952 — the clean-sweep message names BOTH unprovable residual classes; a surviving-but-empty-hash dead-letter row is correctly NOT convicted =="
-(
-  WAKE_STATE_HOME="$(fresh_state t17)"
-  export WAKE_STATE_HOME
-  unset WAKE_AGENT
-  STATE_DIR="$WAKE_STATE_HOME/default"
-  rec="$STATE_DIR/consumed-hashes.jsonl"
-  dl="$STATE_DIR/dead-letter.jsonl"
-  # Drive the REAL flow so the consumed-hashes row is produced by the actual
-  # writer (_record_last_consumed), not hand-built: 12 fillers, then the bench
-  # key lands at observed_seq 13 — the same seq as the live specimen below.
-  i=1
-  while [ "$i" -le 12 ]; do
-    "$STORE" enqueue --class actionable --locators "{\"kind\":\"filler\",\"id\":\"f$i\",\"observed_hash\":\"HF$i\"}" >/dev/null || fail_msg "T17: filler enqueue $i failed"
-    i=$((i + 1))
-  done
-  "$STORE" enqueue --class actionable --locators '{"kind":"bench","id":"malformed-locator-test","observed_hash":"H-REAL-CONSUMED"}' >/dev/null
-  "$STORE" consume --upto 13 >/dev/null 2>&1 || fail_msg "T17: baseline consume failed"
-  jq_any "$rec" '.kind=="bench" and .id=="malformed-locator-test" and .observed_seq==13 and .observed_hash=="H-REAL-CONSUMED"' ||
-    fail_msg "T17: fixture broken — the real writer did not record the bench@13 row"
-  # VERBATIM live specimen: mos-dt lane dead-letter.jsonl line 1 (2026-07-26).
-  # Copied byte-for-byte per the #952 fixture constraint — nested .locators.*
-  # with NO observed_hash key, so the audit's extraction yields "".
-  printf '%s\n' '{"observed_seq":13,"locators":{"kind":"bench","id":"malformed-locator-test","note":"deliberately non-conformant"},"class":"actionable","emit_ts":1785055610,"hmac":""}' >"$dl"
-  # Guard the near-miss trap: the specimen must be genuinely NESTED (no
-  # top-level kind) and must extract an EMPTY hash via the audit's own paths.
-  jq -e '(has("kind") | not) and .locators.kind=="bench" and ((.locators.observed_hash // "") == "")' "$dl" >/dev/null ||
-    fail_msg "T17: specimen fixture is not the genuine nested shape — rebuild it from the live lane, not the schema"
-  # Evidence SURVIVES (kind/id/seq all match the row) but extracts hash "";
-  # _record_last_consumed never writes an empty-hash row, so the four-field
-  # match can never fire: this is unprovable class 2, NOT a false witness.
-  rep="$TMP_ROOT/t17.rep"
-  "$STORE" quarantine-audit >"$rep" 2>&1 || fail_msg "T17: the audit must exit 0 — nothing here is provable [$(cat "$rep")]"
-  grep -q 'FALSE WITNESS' "$rep" && fail_msg "T17: the empty-hash evidence must NOT convict (the predicate is correct and must not change) [$(cat "$rep")]"
-  # The wording under test (#952): BOTH residual classes, named.
-  grep -qi 'pruned' "$rep" || fail_msg "T17: clean sweep must name residual class 1 — evidence pruned away [$(cat "$rep")]"
-  grep -qi 'empty observed_hash' "$rep" || fail_msg "T17: clean sweep must name residual class 2 — surviving evidence with an empty observed_hash [$(cat "$rep")]"
-  # --repair on a clean sweep removes nothing: the unprovable row survives.
-  "$STORE" quarantine-audit --repair >"$TMP_ROOT/t17.fix" 2>&1 || fail_msg "T17: --repair on a clean sweep must exit 0 [$(cat "$TMP_ROOT/t17.fix")]"
-  jq_any "$rec" '.kind=="bench" and .observed_seq==13 and .observed_hash=="H-REAL-CONSUMED"' ||
-    fail_msg "T17: --repair must NOT remove the unprovable bench@13 row — the audit only removes what the ledger can convict"
-  [ "$(grep -c . "$dl")" = "1" ] || fail_msg "T17: the dead-letter LEDGER must be untouched"
-) && ok
-
 echo
 if [ -s "$FAILFILE" ]; then
   echo "wake store/ack harness: FAILED ($(grep -c . "$FAILFILE") assertion(s))" >&2