Six additional Prettier issue checks

Registered before these task calls and after prior batches and timeouts were observed. The six issues were open with complete comment snapshots at selection. This is another separate instrumentation batch, not a pre-registered 30–50-task cohort. The manifest has SHA-256 3b89ecfbd0610ec073bf826ae3d5d0f2f318ab635cc08e1b6c083ab924895eae and fixes issue snapshots, checks, models and paired random order.

Issue	Frozen issue criterion	Positive reference used only to validate the check
#20174	Both reported TypeScript comment examples produce the stated stable output and preserve the normalized program.	Source hunks of PR #20213, head b48e15cde58e74418d4eea7f6281d1907977fe03
#15537	The reported JavaScript expression is idempotent, retains both comments and preserves the normalized program.	Source hunks of PR #20080, head 545827effc19998655f1363546572fc52186cca1
#17213	The reported TypeScript loop is idempotent, retains the directive immediately before the following expression line, and preserves the normalized program.	Source hunks of PR #19999, head f2c3d19b86c1aec020e5daa5699ae7e7ecd651ba
#18066	The reported Markdown list/comment example retains its stated formatting and is idempotent.	Source hunks of PR #18883, head b57ddf9558c42ed05b976ea7949985b2238e5a43
#9955	The GraphQL example is idempotent, preserves its parsed document and retains the field comment before the selection set.	The source-only port of PR #14838, head ed986c4fcded561e6a0e28b20f59730611d13c48
#9112	The first reported GraphQL query is idempotent, preserves its parsed document and retains both argument comments before the closing parenthesis.	The same source-only port of PR #14838

Each frozen issue/public check fails on pinned Prettier commit bed4edbb1cf2a9db4004215b0ce5707cdb02b5d9 and passes on its source-only reference; two small ordinary-formatting controls pass on both. These narrow tests are not full project acceptance. The normalized program checks omit positions and comment attachment. In particular, #17213 does not run the external ESLint rule or prove that the directive suppresses the intended diagnostic. The three GraphQL issues #8182, #9955 and #9112 share a reference patch and may share a cause; they must not be treated as independent samples.

The restricted VM, 12-turn budget and two randomized pre-seeded contexts follow the alternate-route protocol. For #20174, #17213 and #9112 both retrieval conditions select the same single comment; these are not long-session memory challenges. Public issue discussions and PRs can disclose proposed solutions.

The solver is gpt-6.1-sol, the blinded AI reviewers are deepseek-flash and gpt-6-astra, and the AI adjudicator is claude-haiku-4-5, a third model name chosen after observing the previous adjudication timeout. Model provenance through the proxy is not independently verified and none is a human reviewer. Two earlier Haiku reviews returned empty content; invalid or missing outputs here remain missing. Explicit scoring direction, per-stage stopping and no automatic retry remain as in the prior protocol. The estimated target is below $2, not a hard cap. These tasks do not establish production correctness, savings or autonomous memory management.