fix(sdk): judge fenced llm replies by their content; resume past a predicate gate

docs/SURFACE.md

@@ -399,7 +399,11 @@ authentication probes, and exact `flows.json` model allow-list as agent steps;
 a declared `model` must be in that project's `models` array. A template call
 such as ``await f.llm`Summarize ${text}` `` returns text. The structured overload
 parses JSON and checks `output` before submitting a successful completion;
-the kernel independently checks the schema before accepting the output.
+the kernel independently checks the schema before accepting the output. A
+reply that is exactly one markdown code fence around a value (three or more
+backticks or tildes, closed by a run of the same character at least as long)
+is judged by the value inside it; prose around the JSON, or two fenced values,
+is still invalid.
 Invalid JSON or a schema mismatch completes with `verification_failed` and
 prevents downstream work. Retry and lease handling use the existing kernel
 policies; this overload introduces no separate retry contract.

evidence/llm-fence-predicate-resume/mutation-1-fence.txt

@@ -0,0 +1,58 @@
+$ cp src/llm-worker.ts /var/folders/6d/0x5fkt8d01gfmmjdzkxqzwnh0000gn/T/tmp.oIUGoRyujb/llm-worker.ts
+exit=0
+$ sed -i.bak 's/JSON.parse(unfenced(result.stdout_tail))/JSON.parse(result.stdout_tail)/' src/llm-worker.ts && rm src/llm-worker.ts.bak
+exit=0
+$ git diff --stat -- src/llm-worker.ts
+ packages/sdk/src/llm-worker.ts | 2 +-
+ 1 file changed, 1 insertion(+), 1 deletion(-)
+exit=0
+$ npx vitest run tests/worker-transcript.test.ts -t 'fenced reply'
+
+ RUN  v2.1.9 /Users/khaliqgant/Projects/AgentWorkforce/flows/packages/sdk
+
+ ❯ tests/worker-transcript.test.ts (8 tests | 1 failed | 7 skipped) 205ms
+   × the llm worker judges the value inside one markdown fence > accepts a fenced reply whose content matches the schema, and journals the parsed value 205ms
+     → expected 'verification_failed' to be 'success' // Object.is equality
+
+⎯⎯⎯⎯⎯⎯⎯ Failed Tests 1 ⎯⎯⎯⎯⎯⎯⎯
+
+ FAIL  tests/worker-transcript.test.ts > the llm worker judges the value inside one markdown fence > accepts a fenced reply whose content matches the schema, and journals the parsed value
+AssertionError: expected 'verification_failed' to be 'success' // Object.is equality
+
+Expected: "success"
+Received: "verification_failed"
+
+ ❯ tests/worker-transcript.test.ts:215:27
+    213|   it('accepts a fenced reply whose content matches the schema, and jou…
+    214|     const completion = await completeWith('```json\n{"x": 4}\n```');
+    215|     expect(completion[4]).toBe('success');
+       |                           ^
+    216|     expect((completion[5] as { output: unknown }).output).toEqual({ x:…
+    217|   });
+
+⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯[1/1]⎯
+
+ Test Files  1 failed (1)
+      Tests  1 failed | 7 skipped (8)
+   Start at  22:18:46
+   Duration  1.36s (transform 256ms, setup 0ms, collect 901ms, tests 205ms, environment 0ms, prepare 49ms)
+
+exit=1
+$ cp /var/folders/6d/0x5fkt8d01gfmmjdzkxqzwnh0000gn/T/tmp.oIUGoRyujb/llm-worker.ts src/llm-worker.ts
+exit=0
+$ cmp src/llm-worker.ts /var/folders/6d/0x5fkt8d01gfmmjdzkxqzwnh0000gn/T/tmp.oIUGoRyujb/llm-worker.ts
+exit=0
+$ git diff --stat -- src/llm-worker.ts
+exit=0
+$ npx vitest run tests/worker-transcript.test.ts -t 'fenced reply'
+
+ RUN  v2.1.9 /Users/khaliqgant/Projects/AgentWorkforce/flows/packages/sdk
+
+ ✓ tests/worker-transcript.test.ts (8 tests | 7 skipped) 195ms
+
+ Test Files  1 passed (1)
+      Tests  1 passed | 7 skipped (8)
+   Start at  22:18:48
+   Duration  901ms (transform 206ms, setup 0ms, collect 461ms, tests 195ms, environment 0ms, prepare 43ms)
+
+exit=0

evidence/llm-fence-predicate-resume/mutation-2-predicate.txt

@@ -0,0 +1,58 @@
+$ cp src/authored-flow-executor.ts /var/folders/6d/0x5fkt8d01gfmmjdzkxqzwnh0000gn/T/tmp.oIUGoRyujb/authored-flow-executor.ts
+exit=0
+$ sed -i.bak 's/JSON.stringify(canonical)/JSON.stringify(record)/' src/authored-flow-executor.ts && rm src/authored-flow-executor.ts.bak
+exit=0
+$ git diff --stat -- src/authored-flow-executor.ts
+ packages/sdk/src/authored-flow-executor.ts | 2 +-
+ 1 file changed, 1 insertion(+), 1 deletion(-)
+exit=0
+$ npx vitest run tests/authored-agent-artifacts.test.ts -t 'sorted keys'
+
+ RUN  v2.1.9 /Users/khaliqgant/Projects/AgentWorkforce/flows/packages/sdk
+
+ ❯ tests/authored-agent-artifacts.test.ts (5 tests | 1 failed | 4 skipped) 42ms
+   × predicate verdicts recorded on the root run > a resumed verdict read back with sorted keys lowers the same gate command as the first run 41ms
+     → expected 'printf \'%s\' \'{"because":"plan cove…' to be 'printf \'%s\' \'{"gate":"predicate","…' // Object.is equality
+
+⎯⎯⎯⎯⎯⎯⎯ Failed Tests 1 ⎯⎯⎯⎯⎯⎯⎯
+
+ FAIL  tests/authored-agent-artifacts.test.ts > predicate verdicts recorded on the root run > a resumed verdict read back with sorted keys lowers the same gate command as the first run
+AssertionError: expected 'printf \'%s\' \'{"because":"plan cove…' to be 'printf \'%s\' \'{"gate":"predicate","…' // Object.is equality
+
+Expected: "printf '%s' '{"gate":"predicate","step":"run-1","verdict":"pass","because":"plan covers every question"}'"
+Received: "printf '%s' '{"because":"plan covers every question","gate":"predicate","step":"run-1","verdict":"pass"}'"
+
+ ❯ tests/authored-agent-artifacts.test.ts:300:25
+    298|     expect(stream).toHaveLength(1);
+    299|     expect(commands).toHaveLength(2);
+    300|     expect(commands[1]).toBe(commands[0]);
+       |                         ^
+    301|     expect(commands[0]).toContain('{"gate":"predicate","step":"run-1",…
+    302|   });
+
+⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯[1/1]⎯
+
+ Test Files  1 failed (1)
+      Tests  1 failed | 4 skipped (5)
+   Start at  22:18:50
+   Duration  1.56s (transform 573ms, setup 0ms, collect 1.26s, tests 42ms, environment 0ms, prepare 66ms)
+
+exit=1
+$ cp /var/folders/6d/0x5fkt8d01gfmmjdzkxqzwnh0000gn/T/tmp.oIUGoRyujb/authored-flow-executor.ts src/authored-flow-executor.ts
+exit=0
+$ cmp src/authored-flow-executor.ts /var/folders/6d/0x5fkt8d01gfmmjdzkxqzwnh0000gn/T/tmp.oIUGoRyujb/authored-flow-executor.ts
+exit=0
+$ git diff --stat -- src/authored-flow-executor.ts
+exit=0
+$ npx vitest run tests/authored-agent-artifacts.test.ts -t 'sorted keys'
+
+ RUN  v2.1.9 /Users/khaliqgant/Projects/AgentWorkforce/flows/packages/sdk
+
+ ✓ tests/authored-agent-artifacts.test.ts (5 tests | 4 skipped) 33ms
+
+ Test Files  1 passed (1)
+      Tests  1 passed | 4 skipped (5)
+   Start at  22:18:52
+   Duration  982ms (transform 371ms, setup 0ms, collect 736ms, tests 33ms, environment 0ms, prepare 57ms)
+
+exit=0