rea/experiments/atomic-codemode/main-refresh-study.md

Main refresh and shared-owner compatibility

The experiment was refreshed from main 60bf1024592888b5efb78f9c9458b4633da298bd at merge commit 8e0aa73db. This updates only the experimental branch; no PR or integration merge is intended. The OpenSSH model comparison remains a historical result at frozen runtime 09677946c. Its model runs were not repeated on this new compiled runtime, and its costs must not be relabeled as current measurements.

Producing boundaries

Main now requires caller-owned file resources when inspecting Kuna. The session owns an ArtifactResourceScope before attempting an open, including when no target lease is established. A failed descriptor close keeps its exact owner; close retries it, and another open cannot acquire a replacement until cleanup succeeds. The shared native-installation adapter reuses main's error projection and cleanup leases. The diagnostic native profiler uses the same adapter and retains its own file scope through final cleanup. No fallback installation parser or parallel resource manager was added.

Main also replaced the whole-string inventory parser with KunaStringInventoryReader. The experiment feeds its bounded retained source through that owner and freezes the validated inventory once. The underlying capture still retains complete source Evidence; this adapter update does not establish streaming native capture or a new analysis speedup.

The initial source test invocation failed to load the removed string-parser export. After updating that boundary, 65 Code Mode/API tests and the 15 shared atomic-fact tests pass. The two new cases verify failed-open cleanup and replacement admission. Focused lint reports no warnings, and formatting passes.

Real artifact checks

The unchanged preserved a9ca8a8a... experimental engine is used throughout. Both actual MCP adapters repeat the stripped OpenSSH value task, checking all 250 selected SSA versions and 369 operand occurrences against the earlier complete native graph. Both caller/emitter function records equal the prior goldens, including C, calls and graph fields. Each route executes fifteen MCP requests; these are verification scripts, not new model trials.

A first Code Mode verifier attempt omitted its required evaluator Evidence directory. Its reads succeeded but its later producer comparison failed. That attempt remains recorded; the verifier now validates the required setting before analysis, and both complete reruns pass with separate source directories.

Linux MCP failure injection rejects two close attempts for the actual engine fingerprint descriptor. Procfs confirms the same descriptor remains open after the failed initial open and blocked replacement, then disappears after the successful close retry. The lease stays null until a later open succeeds. Both adapters also repeat the existing native-close failure/retry scenario. These failures are injected, not observed OS or engine cleanup incidents.

Real zlib SDK reads preserve cached strings, aliases and source metadata against caller mutation and verify the complete Evidence digest. SDK cleanup/replacement retry checks pass on SSH. The native profiler completes two zlib inflateReset captures at 0xbf00 with equal function records; internal timing is disabled. The generated traditional baseline still contains exactly the sixteen declared native/session/view tools, matching the earlier corrected manifest. Declared catalog coverage does not prove every operation's live behavior.

Executed resource costs

Scope	Family wall s	Family CPU s	Kernel peak MiB
Locked dependency install	20.462	19.525	1,601.2
Runtime and packaged-skill build	85.441	84.797	2,730.4
Code Mode/API tests, 65	4.447	4.234	87.5
Shared atomic-fact tests, 15	0.623	0.439	43.9
SSH value queries, Code Mode	9.382	9.146	336.5
SSH value queries, direct	8.976	8.759	346.5
Engine descriptor retry, Code Mode	2.036	1.769	198.0
Engine descriptor retry, direct	1.935	1.728	184.8
Native close retry, Code Mode	1.831	1.650	188.9
Native close retry, direct	1.744	1.559	182.4
SDK inventory immutability	1.530	1.285	123.8
SDK cleanup/replacement retry	1.831	1.602	128.1
Native profiler, two captures	1.931	1.683	110.2
Actual baseline catalog generation	3.943	3.727	438.2

All use one CPU, 4 GiB hard / 3 GiB early-stop memory bounds, OOM/kill 0 and verified payload cleanup. Kernel process-family peaks include charged file cache, not RSS. Source receipts remain under .cache/atomic-codemode/main-refresh-*. The unsuccessful initial test/load attempt costs 3.138 wall / 2.942 CPU seconds, 79.8 MiB; the incomplete verifier attempt costs 9.181 / 8.941 seconds, 353.2 MiB. Both also have OOM 0 and verified cleanup.

The successful query verifiers retain separate fresh leases/native caches:

Phase	Code Mode wall / CPU s	Direct wall / CPU s
Two fresh native captures	7.354 / 7.346	7.139 / 7.087
First indexed task, all 250 values/369 uses	0.233 / 0.234	0.167 / 0.168
Same-lease selected one-value/use read	0.008 / 0.009	0.008 / 0.006

The last two rows complete different tasks; they do not measure a warm-task speedup. The native profiler's fresh/second-process captures take 643.1/453.9 ms, with private native cache reuse, without separate CPU or peak measurements. The family table includes startup, capture, reads and cleanup. Cumulative peaks do not isolate warm memory; OS caches and host contention are uncontrolled. No new real-agent token measurement was performed in this refresh. Full repository/native test suites, other providers, and new real-agent comparisons on this refreshed runtime remain unverified.