fix(test): identify pi's compaction summarization without sniffing its prompt wording

[milanagm]

Why

main has been red since 2026-09-20 in the Test (pi latest, …) matrix cell only — the 0.83.0, 0.84.1 and devDependency cells, Typecheck and the tarball job all pass (last green nightly: 09-19). Every open PR inherited the same two failures, so each one looked self-inflicted:

• traces the compaction summarization call so the session total matches pi — 0 !== 1
• carries the compaction summary after a compaction, not the messages it replaced

This is a stale test fixture, not a plugin defect. No src/ change: the plugin traced correctly throughout.

Root cause

The mock provider identified pi's compaction-summarization request by the <conversation> XML wrapper around the history:

if (JSON.stringify(messages).includes("<conversation>")) { /* answer with a summary */ }

pi 0.87.1 (earendil-works/pi#9908, split-turn summaries being refused by Claude Fable 5.1) rewrote the turn-prefix summarization prompt to markdown headings. The main history path still uses <conversation>, which is why 0.86.0 was unaffected and 0.87.1 is not:

pi 0.86.0  compaction.js:653  `<conversation>\n${conversationText}\n</conversation>\n\n${TURN_PREFIX_…}`
pi 0.87.1  compaction.js:751  `# Conversation\n${conversationText}\n\n# Instructions\n${TURN_PREFIX_…}`

The summarization call still happened — the mock just stopped recognizing it, fell through to the ordinary stage-based agent script, and answered a summarization request with a tool call. Compaction could then never complete, so nothing was traced. Measured on unmodified main, using tools.length === 0 as ground truth for "this is a summarization call":

pi	real summarization calls	old detector fires	Compaction spans
0.85.1	req#3	yes	1 → green
0.87.1	req#2, req#4	no	0 → 0 !== 1

Note this is invisible in the CI log: the suite prints TAP output only, never provider request bodies.

What changed

1. Identify summarization structurally, not by wording. A summarization call reaches the provider through completeSummarization, which bypasses the agent loop and therefore declares no tools; every agent-loop call offers the built-in tools. Verified stable across pi 0.83 → 0.87.
2. Guard the mock's script against the post-compaction rewind. stage counts visible tool results, and a compaction replaces the ones it summarized — so the script rewound, re-served the read README.md it had just summarized away, grew the context back over the threshold, and pi compacted again without end (11k+ provider calls until the harness SIGKILL). The guard keys on the mock's own summary text, not pi's.
3. Replace the hard-coded compaction count with one span per summarization call. pi ≤0.86 sizes the context from the usage the provider reported (~1678 tok here), pi 0.87 from its own estimate (~4058 tok), so the same sandbox legitimately compacts once or twice and no single threshold value can satisfy both. One-span-per-call is the invariant this plugin actually owes, and it still catches double-tracing.

Verification

Full suite green on every matrix version, including 0.86.0, which previously livelocked:

pi	0.83.0	0.84.3 (devDep)	0.85.1	0.86.0	0.87.1 (latest)
result	128/128	128/128	128/128	128/128	128/128

Typecheck clean. Also verified end-to-end by applying this patch to two open PRs and running against latest: #41 → 141/141, #46 → 143/143 — #46's 143 is exactly the total CI reported as "143 tests, 141 pass, 2 fail".