fix: revive the tp-* response-budget watchdog (0.49.0)

[Shahinyanm]

What

Revives the tp-* response-budget watchdog, which has been dead for months, and cuts 0.49.0.

Why

Every tp-* agent declares a Response budget: ~N tokens, and a watchdog was supposed to flag replies that overran it. It never ran once.

The check lived in PostToolUse:Task, which does not fire for the dispatch tool on current Claude Code. This repo already knew that — installer.ts has said so since v0.39.3 — but nobody connected it to the budget path. The evidence was quiet and total: .token-pilot/over-budget.log did not exist on any machine, and every recorded task event carried a null budget.

It surfaced while chasing a suspected undercount in token accounting. There was no undercount, but this was sitting next to it.

How

• Moved the check to SubagentStop, where subagent completions actually arrive.
• Corrected what it measures. It now sizes the agent's final reply rather than everything the agent emitted along the way. That distinction decides whether the feature is usable at all: summing all output gives ~55k against a 600-token budget, so every agent would be flagged forever.
• The source is the transcript, not the payload's last_assistant_message. That field is present in the 2.1.220 bundle, but nothing in this codebase has ever read it, so there is no evidence it gets populated. The transcript is a path we already depend on and verified against real data.
• Reuses the existing loadAgentBody / parseAgentBudget / decideBudgetAdvice / appendOverBudgetLog helpers rather than reimplementing them.
• over-budget.log is written unconditionally. Surfacing the warning in the transcript rides the existing TOKEN_PILOT_SUBAGENT_FEEDBACK gate, because additionalContext on SubagentStop needs CC 2.1.163+.

Risk

Low. checkSubagentBudget returns null on every negative path — non-tp agent, no budget declared, agent file missing, transcript unreadable — and is wrapped so it cannot throw. A watchdog must not be able to break subagent completion.

Verified end to end on a built dist: an overrun writes agent=tp-debugger budget=100 actual=2000 over=19.00, and a reply within budget writes nothing while still recording its task event. 11 new tests, full suite 1438 green.

Notes for reviewer

The measurement axis is the judgement call worth checking. "Response budget" reads as the size of the answer handed back, so that is what this measures. If the intent was the agent's total output instead, the threshold in every agent file is wrong by two orders of magnitude and the fix belongs there rather than here.