fix: resolve 5 remaining pre-publish blockers (14, 15, 17, 21, 25c)

- completion-promise-detector: restrict to assistant text parts only,
  remove tool_result from completion detection (blocker 14)
- ralph-loop tests: flip tool_result completion expectations to negative
  coverage, add false-positive rejection tests (blocker 15)
- skill tools: merge nativeSkills into initial cachedDescription
  synchronously before any execute() call (blocker 17)
- skill tools test: add assertion for initial description including
  native skills before execute() (blocker 25c)
- docs: sync all 4 fallback-chain docs with model-requirements.ts
  runtime source of truth (blocker 21)

Verified: bun test (4599 pass / 0 fail), tsc --noEmit clean

src/hooks/ralph-loop/completion-promise-detector.ts

@@ -63,6 +63,7 @@ export function detectCompletionInTranscript(
 			try {
 				const entry = JSON.parse(line) as TranscriptEntry
 				if (entry.type === "user") continue
+				if (entry.type !== "assistant" && entry.type !== "text") continue
 				if (startedAt && entry.timestamp && entry.timestamp < startedAt) continue
 				const entryText = extractTranscriptEntryText(entry)
 				if (!entryText) continue
@@ -128,7 +129,7 @@ export async function detectCompletionInSessionMessages(
 
 			let responseText = ""
 			for (const part of assistant.parts) {
-				if (part.type !== "text" && part.type !== "tool_result") continue
+				if (part.type !== "text") continue
 				responseText += `${responseText ? "\n" : ""}${part.text ?? ""}`
 			}