fix(own-loop): prune fires on cumulative tokens; compaction logs truthfully

[jonnyparris]

What

Two related fixes found by deploying #75 and watching wrangler tail on a Gemma session.

1. Prune trigger considers cumulative tokens (the actual failure mode)

PR #75 gated the prune on projectedInputTokens >= threshold — the size of the next outbound message array. For typical Gemma sessions each per-step call is ~19k tokens (system prompt + tool defs dominate), so the per-iteration projection stays small even as the cumulative climb burns through the budget across many iterations.

Empirical reproduction (8 successful steps observed, no prune log fired):

step 0  cumulative=19569   budget=10%   tools=[explore]
step 1  cumulative=39208   budget=19%
step 2  cumulative=58897   budget=29%
step 3  cumulative=78665   budget=38%
step 4  cumulative=98512   budget=48%
step 5  cumulative=118492  budget=58%   ← mid-loop trigger fires
step 6  cumulative=138061  budget=67%
step 7  cumulative=157700  budget=77%   ← wrap-up warning fires
step 8: ...model bails

Now triggers on EITHER signal. Log gains trigger: cumulative | projected.

2. Compaction logs distinguish 'summarised' vs 'noop'

The own-loop's three compaction trigger sites (loop-entry, mid-loop, pre-step) all logged compaction triggered unconditionally — even when maybeCompactContext() no-op'd because Think's persisted history had nothing to summarise. That made the bug invisible. We hit this twice already (see PR #74's caveat).

Now each site:

1. Reads getCompactionCount() before the attempt
2. Calls maybeCompactContext
3. Reads again, logs outcome: "summarised" | "noop"

Log name also changed from *compaction triggered → *compaction attempted for honesty.

New private helper getCompactionCount() keeps the call sites clean.

Tests

All 28 existing compaction + prune tests pass. Typecheck clean. No new tests — both changes are observability + trigger-condition fixes, not new algorithms.

Linked

• PR fix(compaction): force/truncated bypass the minMessages guard #73, fix(compaction): mid-loop trigger uses force=true; guard slice underflow #74, feat(own-loop): in-memory tool-result prune harness safety net #75 (compaction-reachability + harness prune chain)
• memory/learnings/dodo-orchestrator-model-matters.md

beep-boop-🤖