Decompose spawnTuiAgent() into cohesive helpers (#2833)

[atomantic]

Summary

spawnTuiAgent() in server/services/agentTuiSpawning.js had grown to ~1000 lines, handling PTY spawn, output buffering/spooling, sentinel polling, idle/runtime timeouts, paste-marker handshaking, and failure analysis in one function body. This extracts the two most cohesive, self-contained sub-concerns into sibling modules so spawnTuiAgent reads as an orchestrator:

• agentTuiSpawning/outputSpooler.js — createOutputSpooler({ agentId, outputFile, rawFile }) owns both batched write pipelines: parsed status lines → output.txt + agent state stream, and raw PTY bytes → the raw.txt disk spool (with the once-per-run overflow warn + metadata flag, the disk-safety-valve truncation, and the 250ms debounce). Exposes appendLine, pushRaw, flushRaw, drainLines, drainRaw, getOutputBuffer.
• agentTuiSpawning/finalizeHelpers.js — the failure-analysis + worktree-inspection helpers that were module-level functions: readFileTail, worktreeHasChanges, captureWorktreeDiff, plus a new resolveErrorAnalysis that folds the finalize-path "read the raw-spool tail, else fall back to the output buffer, else honor an immediate-fallback signal" logic into one call. RAW_TAIL_ANALYSIS_BYTES moves here.

The main module drops from ~1385 to ~1166 lines. No behavior change to spawn/timeout/completion semantics — the extracted code is byte-for-byte the same logic, only relocated, and every call site in spawnTuiAgent (finish() drain + error-analysis + outputBuffer reads, handleData raw push, finishStartupFailure raw flush) is rewired to the new surface with identical ordering.

Test plan

• npx vitest run services/agentTuiSpawning.test.js — all 54 existing agent-spawn tests pass unchanged (idle-complete/idle-no-activity/idle-no-changes, max-runtime, paste handshake, MCP-boot budget, raw-spool truncation, tail-read, sentinel ingestion, etc.).
• New services/agentTuiSpawning/outputSpooler.test.js (7 cases) covers appendLine flush + dedup + blank-drop, output-buffer overflow warn/metadata (once), raw append vs. truncation dispatch, and no-op drains.
• New services/agentTuiSpawning/finalizeHelpers.test.js (15 cases) covers readFileTail (missing → null, empty → '', whole-file, tail-slice), worktreeHasChanges (dirty/clean/reject/empty-path), captureWorktreeDiff (combined write + return, no-diff, invalid args), and resolveErrorAnalysis (success skips read, immediate-fallback short-circuit, tail analyze, missing-spool fallback).
• npx vitest run services/agentManagement.test.js — 34 pass (downstream consumer unaffected).

Closes #2833