ci: add copilot code-review instructions

[juan-malbeclabs]

Summary of Changes

• Adds .github/copilot-instructions.md (repo-wide review posture + cross-cutting rules) and eight path-scoped .github/instructions/*.instructions.md files, so Copilot's automated PR review applies the standards we already enforce by hand instead of the generic ones.
• Scoped by applyTo glob so each review only loads what the diff touches: onchain programs, Rust SDK/builders/CLI, Go, E2E/QA, release workflows, the language SDKs and their golden fixtures, tests, and the changelog.
• No effect on builds or runtime — these files are only read by Copilot.

Where the rules come from

1. Our own reviews. I pulled every inline review comment on PRs merged in the last month — 421 comments across PRs controller: write config agent version info to ClickHouse #3578–serviceability: address review comments on dzf_locked set-flags #4089 — and grouped the recurring findings. The threshold for a rule was that the same pattern appeared in two or more distinct PRs, or is already codified in an RFC or docs/cli-standard.md. Both sides of each thread were read, so suggestions that were deliberately declined (per-environment payload formats, the grace-period bound, hoisting at three copies) did not become rules.
2. Published RFCs. RFC-1 contributes semver classification and the requirement to state deploy-ordering prerequisites; RFC-10 contributes the compatibility-window schema-evolution invariant (read both formats, keep writing the old one until min_compatible_version clears every reader); RFC-20 gates new CLI verbs.
3. The code itself. Every factual claim in the rules — symbol locations, SDK_ATTACHES_PERMISSION_ACCOUNTS, the absent overflow-checks profile, split_trailing_permission semantics, the goccy/go-json decode path — was verified against the source rather than taken from the review text.

Rules were then pruned to those where a miss ships a real defect: silent feature disable, broadened authorization, a wire break, an unreclaimable reference count, a release tagged against the wrong commit, an alert that never fires. Style, naming, redundancy-removal and log-clarity nits were dropped, since noisy automated review is worse than none.

Notable gap this closes: nothing previously told an automated reviewer how to treat a pre-existing defect the diff widens — a framing that appears in six of the PRs sampled and shapes most human comments in the sample.

Testing Verification

• Verified every applyTo glob matches tracked files — no dead patterns.
• Replayed the file lists of five recent PRs (client/liveness: bound route reconvergence after transient loss (#3935) #4086, serviceability: add dzf_locked access-pass flag and set-flags instruction #4083, ci: testnet release post-dry-run fixes (single tag approval, pipeline-complete guard) #4041, client: add multicast subscriptions to doublezero status output #3964, serviceability: change user feed_pk to vec of feed_pks #4080) to confirm each activates the intended instruction set, which caught two gaps now fixed: client/doublezero/** matched nothing, and the Go serviceability decoder lives at smartcontract/sdk/go/, not under sdk/.
• Fact-checked the rules against the code; this corrected a wrong file path for AUTHORIZE_GATED_FLAGS/legacy_keys_for_flags/check_legacy_any (they are in authorize.rs, not permission/audit.rs), a stale claim that the SDK still appends the Permission PDA, and a rule that would have flagged the system-program check present in 56 processor files.

Follow-up

Path-specific instructions are honored by Copilot code review only when custom instructions are enabled for the repo/org. Worth confirming on this PR's own Copilot review — it touches CHANGELOG.md, so the changelog rules should visibly apply. If they do not, only the repo-wide file is in effect.