fix: pin legal parallel completion observations without hiding outcomes

Repeat replay exposed swapped adjacent activity completions from independent parallel reads. Extend the existing named completion normalization to adjacent operation_activity_completed only, keyed by established span_id; preserve exact outcomes/errors, span links, count and every start/error/other-event boundary. Regression controls reject wrong outcomes/span links and prove boundaries are not crossed. Recorded real DeepSeek fixture outputs now reach loopback POST /v1/chat/completions and retain reasoning/usage/error/stray-delta facts. Correct inaccurate Plan read-only catalog fixture prose without changing captured bytes.

At ffc6b3dbe4 governed all-feature build completed; test result FAILED:10passed/2failed: actual parallel observation race plus intended DeepSeek golden drift after constructor fix. All7hardening controls passed; three other families matched. Frozen-binary corrected SSE recording1passed/0failed (11cases). Original failing diff/logs preserved under conformance-qualification. Final recompile/repeat/default/negative-control qualification pending; no release/platform/provider claims.

Signed-off-by: CodeWhale Bot <[email protected]>

crates/tui/src/conformance/events.rs

@@ -14,6 +14,8 @@
 //! system prompt and tool catalog bodies are replaced by a marker because the
 //! prompt family owns those bytes; runs of adjacent `tool_call_complete`
 //! events are ordered by tool call id because parallel completions race.
+//! Adjacent `operation_activity_completed` observations are likewise ordered
+//! by their established span id. No start, error, or other event is crossed.
 
 use std::collections::{BTreeMap, VecDeque};
 use std::path::Path;
@@ -453,21 +455,68 @@ pub(super) fn normalize_events(events: &[Event], masker: &mut Masker) -> Vec<Val
 fn order_parallel_completions(lines: &mut [Value]) {
     let mut start = 0;
     while start < lines.len() {
+        let (kind, key) = match lines[start]["event"].as_str() {
+            Some("tool_call_complete") => ("tool_call_complete", "tool_call_id"),
+            Some("operation_activity_completed") => ("operation_activity_completed", "span_id"),
+            _ => {
+                start += 1;
+                continue;
+            }
+        };
         let mut end = start;
-        while end < lines.len() && lines[end]["event"] == "tool_call_complete" {
+        while end < lines.len() && lines[end]["event"] == kind {
             end += 1;
         }
         if end - start > 1 {
-            lines[start..end].sort_by(|left, right| {
-                left["tool_call_id"]
-                    .as_str()
-                    .cmp(&right["tool_call_id"].as_str())
-            });
+            lines[start..end].sort_by(|left, right| left[key].as_str().cmp(&right[key].as_str()));
         }
         start = end.max(start + 1);
     }
 }
 
+#[test]
+fn parallel_completion_projection_keeps_outcomes_spans_and_causal_boundaries() {
+    let first = json!({"event": "operation_activity_completed", "span_id": "call_a#1", "outcome": "succeeded"});
+    let second = json!({"event": "operation_activity_completed", "span_id": "call_b#2", "outcome": "failed"});
+    let boundary = json!({"event": "error", "message": "exact error bytes"});
+    let start = json!({"event": "operation_activity_started", "span_id": "call_a#1"});
+    let original = vec![
+        start.clone(),
+        second.clone(),
+        first.clone(),
+        boundary.clone(),
+        second.clone(),
+        start.clone(),
+        first.clone(),
+    ];
+    let expected = vec![
+        start.clone(),
+        first.clone(),
+        second.clone(),
+        boundary,
+        second,
+        start,
+        first,
+    ];
+    let mut normalized = original.clone();
+    order_parallel_completions(&mut normalized);
+    assert_eq!(normalized, expected);
+    assert_eq!(normalized.len(), original.len());
+
+    let mut wrong_outcome = original.clone();
+    wrong_outcome[1]["outcome"] = json!("succeeded");
+    order_parallel_completions(&mut wrong_outcome);
+    assert_ne!(wrong_outcome, expected, "wrong outcome was normalized away");
+
+    let mut wrong_span = original;
+    wrong_span[1]["span_id"] = json!("call_other#2");
+    order_parallel_completions(&mut wrong_span);
+    assert_ne!(
+        wrong_span, expected,
+        "wrong span relationship was normalized away"
+    );
+}
+
 fn check_invariants(case: &Value, workspace: &Path, provider: &ScriptedProvider) -> Vec<String> {
     let mut violations = Vec::new();
     for invariant in case["invariants"].as_array().into_iter().flatten() {

crates/tui/tests/fixtures/conformance/README.md

@@ -10,7 +10,8 @@ written in the abstract.
 Everything here is language-neutral: JSON, JSONL, raw SSE bytes, and POSIX
 shell. No fixture was captured from a real provider or contains a secret.
 
-The Rust runner is `crates/tui/src/conformance/` (one test per family):
+The Rust runner is `crates/tui/src/conformance/` (one golden test per family,
+plus hardening controls):
 
 ```sh
 cargo test -p codewhale-tui --lib -- conformance::
@@ -78,6 +79,9 @@ Each line is `codewhale_protocol::EventMsg` as serialized (`"event"` tag), with:
 - `tool_catalog` and `system_prompt` bodies → `"<pinned by the prompt family>"`;
 - runs of adjacent `tool_call_complete` events ordered by `tool_call_id`
   (parallel completions race);
+- runs of adjacent `operation_activity_completed` observations ordered by
+  the established `span_id` for the same reason; outcomes, span relationships,
+  and event counts stay exact, and no start/error/other event is crossed;
 - keys sorted.
 
 The last line is `{"harness_summary": {model_requests, non_streaming_requests,
@@ -179,6 +183,9 @@ background observers, cancellation and process-tree teardown are not covered.
 The turns do not qualify persistence/resume, nested multi-tool cardinality,
 non-draining client backpressure, or every client projection. Slash-command,
 script-tool, and migrated adapter calls need their own fixtures before flipping.
+The Plan prefix records its advertised catalog and frozen prompt bytes; it
+does not prove read-only admission. At this baseline it still advertises
+`bash`, `edit` and `write`; full-turn Plan refusal needs separate evidence.
 
 The prefix runner names unavailable host-probed tools explicitly. A skipped
 definition (`code_execution`, `image_ocr`, `js_execution`, `pandoc_convert`) is

crates/tui/tests/fixtures/conformance/prompt/plan_mode.case.json

@@ -1,5 +1,5 @@
 {
-  "description": "Plan mode: the read-only tool surface and the Plan prompt layer.",
+  "description": "Plan mode: pins the advertised catalog and frozen prefix. The current catalog still advertises bash/edit/write; read-only admission and appended Plan instructions are not proved by a prefix fixture.",
   "model": "deepseek-v4-pro",
   "mode": "plan",
   "approval_mode": "suggest",

crates/tui/tests/fixtures/conformance/sse/deepseek_error_frame.golden.jsonl

@@ -1,2 +1,9 @@
-{"request":[]}
-{"type":"open_failure","detail":"DeepSeek model \"deepseek-v4-pro\" uses ChatCompletions, but this client is bound to Responses; resolve a new model route before sending"}
+{"request":[{"method":"POST","path":"/v1/chat/completions"}]}
+{"type":"message_start","message":{"id":"","type":"message","role":"assistant","content":[],"model":"deepseek-v4-pro","stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":0,"output_tokens":0}}}
+{"type":"content_block_start","index":0,"content_block":{"type":"text","text":""}}
+{"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Partial"}}
+{"type":"error","error":{"message":"Model not exist.","type":"invalid_request_error","code":"invalid_request_error"}}
+{"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" after the error"}}
+{"type":"content_block_stop","index":0}
+{"type":"message_delta","delta":{"stop_reason":"stop","stop_sequence":null}}
+{"type":"message_stop"}

crates/tui/tests/fixtures/conformance/sse/deepseek_reasoning_usage.golden.jsonl

@@ -1,2 +1,11 @@
-{"request":[]}
-{"type":"open_failure","detail":"DeepSeek model \"deepseek-v4-pro\" uses ChatCompletions, but this client is bound to Responses; resolve a new model route before sending"}
+{"request":[{"method":"POST","path":"/v1/chat/completions"}]}
+{"type":"message_start","message":{"id":"","type":"message","role":"assistant","content":[],"model":"deepseek-v4-pro","stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":0,"output_tokens":0}}}
+{"type":"content_block_start","index":0,"content_block":{"type":"thinking","thinking":""}}
+{"type":"content_block_delta","index":0,"delta":{"type":"thinking_delta","thinking":"Compare the two numbers. "}}
+{"type":"content_block_delta","index":0,"delta":{"type":"thinking_delta","thinking":"9.11 < 9.9."}}
+{"type":"content_block_stop","index":0}
+{"type":"content_block_start","index":1,"content_block":{"type":"text","text":""}}
+{"type":"content_block_delta","index":1,"delta":{"type":"text_delta","text":"9.9 is larger."}}
+{"type":"content_block_stop","index":1}
+{"type":"message_delta","delta":{"stop_reason":"stop","stop_sequence":null},"usage":{"input_tokens":20,"output_tokens":30,"prompt_cache_hit_tokens":12,"prompt_cache_miss_tokens":8,"reasoning_tokens":22}}
+{"type":"message_stop"}