[mike-thompson-day8]

Fixes the correctness findings on the Pair MCP server, the Story MCP server and the Pair preload's dispatch projection (rf2-lgzco.33, rf2-lgzco.35, rf2-lgzco.47), plus one test-docstring sentence (rf2-i79mi).

What changed

rf2-lgzco.33 (pair-mcp): JSON pred / signals coercion (P2)

• record and watch-until now coerce a JSON-object pred or JSON signals the way watch-epochs coerces its pred. An :app-db signal or a pred :path parses the way get-path parses a path (args/parse-path-arg). A colon-led :equals or :sub element mints a keyword under the id grammar (args/->id-keyword). Any other string stays a string, which is option (a) in the finding. The EDN-string arms are unchanged.
• watch-until now uses the shared record/parse-pred-arg. record's stop object coerces its nested :pred the same way.
• record_test's held-predicate test used to let the stub decide :held?. It now reads the poll form and checks that the keyword path and comparand reach it.
• The review's P3 lead (strip-fns and RegExp) is real. I reproduced it in the fixture's CLJS build: the EDN read-back of registrar-describe threw on a [:re #"…"] schema. strip-fns now turns a RegExp into its source string, and [:re "…"] is the same Malli schema. machine_describe_test.clj evaluates the shipped walk under Babashka, which has no regexp?, so that test now defines one from the JVM predicate.

rf2-lgzco.35 (story-mcp)

• P3: after a run is abandoned at its deadline, read-failures used to read the torn-down frame and answer {:status :pass :total 0}. Each variant now keeps its deadline outcome's records until its next run, and read-failures reads them instead: {:status :error :total 1} with the :rf.error/run-failed record. This is option (a), and it matches the run-variant reply. tools_test.clj's fence test used to pin the :pass; it now expects that error.
• P4, and the ruled pair-mcp mirror: both servers now echo an unknown tool name longer than 128 characters cut to 128 characters plus .... A name more than twice the longest listed name skips the near-name search, since no listed name could fall within its threshold. Each server has its own fix, and tools/mcp-base is untouched.

rf2-lgzco.47 (Pair preload)

• P2, ruled (a): pure.cljc gains cascade-refusal-ops beside the throw set, and the throw set still mirrors Xray's. Three refusals now count: :rf.error/effect-map-shape, a schema failure at :where :event, and a schema failure carrying :rollback? true. Each folds into :errors, which forces :outcome :error and keeps the event out of :no-op?. So dispatch and dispatch-dry-run now return what recipes.md:273 already promises. A schema failure at :sub-return / :fx-args leaves the event committed and is not counted.
• P3: epoch-matches? now matches nothing on a present but unparseable :timing-ms. The watch-epochs tool also refuses such a value up front with :rf.mcp/invalid-arg naming :pred.

rf2-i79mi: the error_boundary_test namespace docstring now says a failed or interrupted nREPL round-trip also reaches relay 1, as invoke-and-guard's docstring does. This is prose only.

Spec, one sentence per change:

• pair-mcp 003-Tool-Catalogue.md: the unknown-tool echo cut, and the :timing-ms refusal.
• story-mcp 001-Wire-Protocol.md: the unknown-tool echo cut.
• story-mcp 002-Tool-Registry.md §read-failures: what it reads after a deadline.

No new tool args, no MCP-base surface, and no descriptor text changed, so tool-descriptors.edn is untouched.

Quality gates

Each exit code below was captured on the same command line as its redirect.

Gate	Baseline (trunk)	With the change
npm --prefix tools/re-frame2-pair-mcp test	exit 0, 421 tests / 1355 assertions	exit 0, 424 / 1371
clojure -M:test in tools/story-mcp	exit 0, 153 / 362	exit 0, 155 / 370
Pair fixture pure-test (shadow-cljs compile pure-test && node target/pure-test.js, run through the lane helper)	exit 0, 56 / 178	exit 0, 59 / 189
skills-structural bb tests (tests/prompts + tests/runtime/*)	—	all 15 files exit 0
scripts/check_doc_slugs.py	—	exit 0 (the spec edits add no links or anchors)

• Pair fixture baseline: taken with the pure.cljc / runtime.cljs edits in place but the trunk test files, so the counts are the trunk's.
• skills-structural: the first run after the strip-fns change caught machine_describe_test.clj at exit 1, which is how the regexp? gap above was found.

Each delta matched the prediction from the diff:

• pair-mcp: +3 tests, +16 assertions.
• story-mcp: +2 tests, +8 assertions.
• fixture: +3 tests, +11 assertions.

Final-base gate: this PR's CI. The local runs were taken before the rebase onto 8fca3b354d, and none of the upstream commits since touch these paths. The live and hermetic MCP job (mcp_live) is left to CI.

Controls (each planted, watched fail, restored)

Each restore was verified by git hash-object against the committed blob, and git diff --quiet exited 0.

Plant	Run exit	What went red	Controls that stayed green
pair-mcp: JSON coercion, unknown-tool echo and near-name guard, and the timing refusal reverted to trunk behaviour	1	10 failures, all in the four new or changed tests	the watch-epochs JSON-pred test; unknown-tool-refused-before-connection
story-mcp: deadline records ignored; echo uncut	1	4 failures	the near-name unknown-tool test; the fence assertions
fixture: refusal fold off, timing check reverted, RegExp passed through	1	6 failures and 1 error	the genuine no-op; the :sub-return / :fx-args non-refusals; the parseable ">=0"

Detail on the reds:

• pair-mcp: with trunk behaviour restored, the unknown-tool refusal for a 32,000-character name was 64,626 characters long. That reproduces the pair-mcp half of the echo finding, which the review had only inspected.
• story-mcp: read-failures read [:pass 0 []], and the refusal was 64,488 characters.