test(server): strengthen regression coverage and harden query-window search

[Graffioh]

Stacked on #709 — this contains only the coverage deltas; the reorganization lives in the refactor PR. Until #709 merges, this PR's diff includes the refactor. Enforcement-only view: Graffioh/lucebox-hub@codex/unit-test-refactor...codex/unit-test-enforcement

Adds model-free regression coverage and tightens existing assertions:

• qwen3/test_query_capture.cpp: table-driven query_capture_slice clipping, disjoint/empty-range rejection, and an exhaustive chunked-capture reconstruction that asserts every query token is written exactly once.
• qwen3/test_qwen3_buffer_plan.cpp / qwen3/test_score_range.cpp: replace single-suite TEST_CASEs with focused cases covering no-KV tail reuse, per-layer legacy scoring state, zero-layer plans, last-forwarded-layer scoring, and the empty-forwarded edge.
• common/test_gguf_mmap.cpp: failed re-open clears the previous mapping, and move construction/assignment transfers ownership; re-open now verifies payload bytes of a differently-sized replacement.
• common/test_restore_delta.cpp: negative prefix rejection and zero-prefix/empty-prompt handling.
• server/test_http_sessions.cpp: every read accessor refreshes session LRU position, reads of missing sessions do not allocate, EMA asserts the literal expected value, and convergence/LRU tests assert exact bounds and turn counts.
• server/test_http_request.cpp: regression for a repeated suffix in an earlier turn being mistaken for the current pflash query.
• server/test_prefix_cache.cpp / server/test_text.cpp: hash_prefix avoids nullptr; disabled-reasoning asserts the parsed reasoning fields.
• server/test_sse_emitter.cpp: restores the stderr content checks in the malformed-tool-call suppression test.

One production change pairs with the new regression test: find_pflash_query_window now bounds the suffix search to max_tokens before search_end, so a repeated phrase in an older turn cannot be mistaken for the current query. Header comment updated to match.

Test plan

• Combined tree (refactor + enforcement) builds clean; test_server_unit discovers 556 cases; ctest passes 554/556 — the only 2 failures are upstream-verbatim MixedMmqPolicy cases segfaulting inside libamdhip64/ggml HIP kernels on this gfx1201+gfx1151 box (pre-existing environment issue).
• The refactor-only base (refactor(test): organize component suites and replace simulated coverage #709) independently builds and passes 541/543 with the same 2 environment failures.

Generated with Devin

[cubic-dev-ai]

1 issue found across 80 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="server/test/server/test_disk_prefix_cache.cpp">

<violation number="1" location="server/test/server/test_disk_prefix_cache.cpp:443">
P2: This test never exercises cold-prefix boundary selection. It builds the cache with a plain MockBackend, whose snapshot_ref() returns no context, so cold_prefix_boundary() always returns 0 at the `!layout_known_` guard; the assertion `result == 0` only checks that guard. The test name and header comment claim it finds the last boundary <= cold_max_tokens, but that logic is never reached. Use MockBackendWithLayout (already defined above) and call cache.learn_layout(0) before cold_prefix_boundary() so layout_known_ is set and the boundary-finding path is actually covered.</violation>
</file>

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[cubic-dev-ai]

P2: This test never exercises cold-prefix boundary selection. It builds the cache with a plain MockBackend, whose snapshot_ref() returns no context, so cold_prefix_boundary() always returns 0 at the !layout_known_ guard; the assertion result == 0 only checks that guard. The test name and header comment claim it finds the last boundary <= cold_max_tokens, but that logic is never reached. Use MockBackendWithLayout (already defined above) and call cache.learn_layout(0) before cold_prefix_boundary() so layout_known_ is set and the boundary-finding path is actually covered.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At server/test/server/test_disk_prefix_cache.cpp, line 443:

<comment>This test never exercises cold-prefix boundary selection. It builds the cache with a plain MockBackend, whose snapshot_ref() returns no context, so cold_prefix_boundary() always returns 0 at the `!layout_known_` guard; the assertion `result == 0` only checks that guard. The test name and header comment claim it finds the last boundary <= cold_max_tokens, but that logic is never reached. Use MockBackendWithLayout (already defined above) and call cache.learn_layout(0) before cold_prefix_boundary() so layout_known_ is set and the boundary-finding path is actually covered.</comment>

<file context>
@@ -0,0 +1,759 @@
+    std::vector<int> boundaries = {1000, 2000, 3000, 4000, 6000, 8000};
+    // Without layout_known_, returns 0.
+    int result = cache.cold_prefix_boundary(prompt, boundaries);
+    TEST_ASSERT(result == 0);  // layout not known yet
+
+    rm_rf(dir);
</file context>