[viztor]

Phase P-H of docs/improvement-plan.md, stacked on #110.

What changed

Item	Change
src/host/patch-method.ts (new)	patchMethod is now the only way the plugin wraps a method on a host object. A dispose restores only while the own descriptor still holds its wrapper; otherwise the wrapper becomes a pass-through to its unpatched method. When the original was inherited, a dispose deletes the own property; when it was own, a dispose redefines its exact descriptor. A same-site reinstall (same id) wraps the true unpatched method, never the earlier install, so either disposal order ends at the exact original. Identity is read from own descriptors, never with === on a property read. 100% per-file coverage floor.
hideResponsesRoute	listProviders, listConfigurableProviders and listModels on llm go through the helper.
decorateModelDiscovery	llm.discoverModels goes through the helper.
decorateModelInfo	llm.resolveModelInfo goes through the helper.
decoratePlaneAdapters	Restructured into one patch per (adapter, method). pi-ai serves deepseek, opencode and opencode-go from one adapter instance, so each patch composes that adapter's plane routes in registry order. This reproduces today's nesting exactly, including the pi-ai-first [INTENDED_CHANGE] characterization row.
rewrapOnAdaptersUpdated	Built and tested, wired to no site. Wiring it into decoratePlaneAdapters would flip the plugin-first [KNOWN_DEFECT] rows, and P-R1 deletes that site.
test/reload.test.ts (new)	apply A, apply B, dispose A, B intact, dispose B, originals restored exactly, for every site, on own and inherited originals. Also runs the reverse disposal order, the real host (tier A), and the whole sequence through apply().
test/patch-method.test.ts (new)	The helper's contract.
test/models-discovery.test.ts	Today's per-site pins (receivers, own-identity restore, restore only while current, the frozen service, the shared-adapter nesting order) stay unchanged. Five new cases, marked "a deliberate change in P-H", record the edge cases on hosts that refuse the patch.
Docs	docs/improvement-plan.md (P-H status and landed record; the global fetch stacking is left for P6), docs/engineering-notes.md (the measured identity trap), AGENTS.md, README.md, CONTRIBUTING.md.

Why: on the real host the three llm-level disposers never restored. ctx.llm is cordis's traceable proxy, and every read of a function-valued property returns a new shadow proxy, so llm.x === decorated was never true. Each reload stacked a wrapper that kept the previous config's decisions, and a disable left them installed. The plane-adapter disposer restored without checking, so disposing A wrote the prototype's method over B.

Out of scope: the globalThis.fetch swap in lifecycle.ts (a global, not a host service, now on P6's work list) and the llmFacade Proxy in responses-provider.ts (a view, not a patch).

Behaviour changes

None. No wrapper's answer changes on a host that accepts the patch. The docs/improvement-plan.md P-H entry records two things. First, three edge cases that the real LlmRuntime and pi-ai adapter cannot reach: a refused listing define, a refused or frozen plane-adapter method, and a reinstall's registeredRoutes. Second, that a reload or a disable on the real host now restores the methods instead of stacking wrappers.

Review

Round	Found	Upheld	Fixed
1	15	15	15
2	9	9	9
3	6	6	6

Round 3's findings were all low severity and were fixed in tests, test comments and the plan of record, with no runtime change. Each new regression probe was checked by mutating the source, watching exactly the new cases fail, and restoring the source byte-for-byte.

Verification

Node	Step	Result
26.12.1	pnpm run check	All 222 files correctly formatted; 0 errors, 18 warnings (limit 21)
26.12.1	vp test run	89 files, 1449 tests passed
26.12.1	shuffle, seeds 1, 42, 1234	89 files, 1449 tests passed, each seed
26.12.1	vp test run --config vitest.mock.config.ts	3 files, 51 tests passed
26.12.1	vp test run --coverage	exit 0; statements 96.71, branches 92.74, functions 96.63, lines 96.7 (floors 95.9/91.3/95.9/95.9); src/host/patch-method.ts at 100/100/100/100
26.12.1	pnpm run build	ESM 2 files, 346.58 kB; build complete
26.12.1	scripts/check.ts	check passed, including "lib/ carries the current build"
24.22.1	shuffle, seed 7	89 files, 1449 tests passed
24.22.1	mock suite	3 files, 51 tests passed
n/a	local-path and credential grep	Grepped the full diff plus untracked files: no local paths and no credential-shaped strings. A canary path and a canary key each matched once, which shows the greps read their input.

Note: the catalog freshness check fails on every branch until the vendor-catalog refresh lands on main. That failure is unrelated to this change.

---

Generated by Claude Code

[viztor]

catalog freshness is red on aec8c1a because the vendor changed its catalog. This PR didn't cause it. The Go plane renamed gemini-4-argon-contributor to gemini-4-argon, added iceman-contributor-free, and its limits page changed with it, so every branch that predates the regenerated tables is stale.

The fix exists. 2ebaffc on #105 takes the regenerated tables unchanged and moves the fixtures and counts with them. d51a3e6 makes the live suite skip free models the gateway has withdrawn. Both are being merged up the stack branch by branch. #106–#110 are syncing now. This branch gets them in the sync pass that follows the review now running on it, and that push re-runs this check.

---

Generated by Claude Code

[viztor]

Second review of P-H

This phase's first reviews ran on the general model because the dedicated reviewer was unavailable at the time. It has now had the dedicated review.

Rounds (lenses: correctness, rules, test matrix)

round	found	upheld	fixed	upheld by lens
1	5	4	4	correctness 1, rules 1, matrix 2
2	3	3	3	correctness 1, matrix 2
3	2	2	2	correctness 1, matrix 1

All findings were low severity.

Fixed

• src/host/patch-method.ts: the rewrapOnAdaptersUpdated JSDoc now says the listener reinstalls the site inside the plugin's own teardown and on every registry change. It counts the reinstalls a release causes, says the caller must dispose it before releasing adapters, and notes that a throw from a reinstall is logged by the host.
• docs/improvement-plan.md: the plan said all three P-H edge changes are unreachable on the real host for one reason. The third one, the live-reload listing, is unreachable for a different reason, and each change now states its own.
• test/reload.test.ts: every site's tier-A reload case now has a probe whose decorated answer must differ from the host's own. Before, the listing and discovery rows only proved the structure, so "survivor still decorates" and "last dispose restores" could not fail. Now they can.
• test/reload.test.ts: the apply()-twice case now requires a patchMethod record on every own function apply() leaves on the stand-in llm and adapter.
• test/architecture.test.ts: a new textual tripwire checks that "patchMethod is the only way to wrap a host method" across everything apply() reaches, with a self-check of planted violations. Over rounds 2 and 3 it was widened to also catch Object.assign/defineProperty/defineProperties/setPrototypeOf, Reflect.setPrototypeOf, and a target reached through a property path (registration?.adapter, this.llm). The self-check now plants 17 violations, up from 13. Nothing new in src/ matches.
• AGENTS.md and docs/improvement-plan.md: they now list the exact four Reflect and four Object methods the rule matches, replacing "any Reflect.* write", and say why a prototype swap is included.

Declined

Verification did not uphold one round-1 finding. Every upheld finding was fixed, and none is left open. There is no runtime behaviour change; the only stricter thing is the architecture tripwire.

The full gate passes: check, unit suite, mock suite, build and bundle gate.

Commit: 3cfacd9

---

Generated by Claude Code