• The active provider in C:\Users\sunda\.cc-switch\cc-switch.db is Zhipu GLM (4f9dea17-b9fd-4aa8-ab90-2222e77b8818). It is still the older Coding Plan Chat configuration: base URL https://open.bigmodel.cn/api/coding/paas/v4, apiFormat=openai_chat, and codexProtocolMode=manual.
• Its persisted codexChatReasoning says supportsThinking=true, thinkingParam=thinking, effortParam=reasoning_effort, and outputFormat=reasoning_content, but the provider meta has no codexReasoningProjection. The model catalog rows for glm-5.3 and glm-5.3-flash also have no schema-v2 reasoning object.
• resolve_codex_chat_reasoning_projection() intentionally fail-closes manual protocol mode: it emits RawReasoningText only for the exact codexReasoningProjection=raw_reasoning_text; any missing/other value becomes ReasoningProjection::None. The streaming converter then accumulates Chat reasoning_content but emits no Responses reasoning SSE when the projection is None.
• Runtime logs prove the route is /responses externally but converts to /chat/completions at https://open.bigmodel.cn/api/coding/paas/v4/chat/completions. The protocol profile for GLM is independently verified as Chat, with readable reasoning from choices[].message/delta.reasoning_content; therefore the upstream did return reasoning and the loss is at CCSM's projection policy boundary.
• Logs also warn that GLM catalog reasoning declarations are ignored because disableAllowed is missing. This is a stale/incomplete persisted catalog issue and explains the warning, but it is not the direct visibility blocker while the explicit provider-level Chat reasoning config remains present.
• The current source preset has already moved Zhipu GLM to native Responses (https://open.bigmodel.cn/api/v1) and includes complete schema-v2 reasoning metadata. The active database provider has not been migrated to that preset, so source and runtime configuration are not the same.

• resolve_codex_chat_reasoning_projection() now preserves the explicit codexReasoningProjection=raw_reasoning_text behavior and, when that field is absent, derives RawReasoningText only from an explicit codexChatReasoning.outputFormat=reasoning_content declaration.
• Manual providers without a reasoning-content declaration still fail closed; reasoning_summary is never synthesized from manual metadata and still requires verified Responses protocol evidence.
• Regression coverage is green: the legacy GLM Chat shape now resolves to raw reasoning, the anti-summary safety test passes, and all 136 codex provider unit tests pass.
• The fix was built and installed with the recovery-capable CCSM transaction. The first attempt intentionally rolled back after detecting the NSIS-patched executable hash; the second attempt used that observed installed hash and succeeded. Final runtime proof: PID 58292, listener 127.0.0.1:15721, /health=200, /status.running=true, listener_role=takeover, and zero failed requests at the verification snapshot.

Do not fix this by only adding disableAllowed to the catalog. For the currently active Chat route, either persist the exact raw projection (codexReasoningProjection=raw_reasoning_text) or move the provider out of manual protocol mode and let a verified, target-bound Chat profile select the projection. A separate migration is needed if the intent is to use the current native Responses preset.