refactor(client): consolidate Subscribe/SubscribeList into shared subscribeCore

[benitogf]

Summary

Closes #144

The single-object and list-glob subscription paths in the Go client were ~280 lines of byte-identical code wrapping a small per-variant decode step. PR #143 itself had to fix the cancel-suppression bug twice in lockstep — the failure mode duplication invites. Now consolidated into one shared core, with each entry point as a thin wrapper that injects the variant-specific decode + delivery.

Behavior is byte-equivalent end-to-end:

• single-object decode: any failure is fatal (terminates the read loop, triggers reconnect)
• list decode: parse failure is fatal; per-item unmarshal failures fire the error callback and continue
• cancel-time error suppression applied at both connect and read failure sites
• cache + retry-count lifecycle preserved exactly as before

Net code change: client/subscribe.go drops 560 → 433 lines.

Test plan

• Full race-enabled test suite green across all packages at 10x sequential iterations locally
• PR fix(client): suppress Subscribe OnError after context cancellation #143's cancel-suppression regression tests (4 subtests for Subscribe/SubscribeList across dial-failure and read-error paths) continue to fail without the production fix and pass with it — verified by stashing the fix on the new code shape
• Multi-round adversarial self-review converged after two consecutive clean passes (rounds 3 and 4 returned zero confirmed findings)
• All five behavioral invariants (single-fatal-decode, list-skip-on-item-error, both-sites-cancel-suppressed, cache-on-error-path, retry-reset-placement) explicitly verified line-by-line against the pre-refactor baseline

This refactor is stacked on top of #143 — when that merges, this PR rebases onto main cleanly.

---

🤖 Generated with Claude Code

[benitogf]

Closing after a tougher review on the headline numbers.

The PR description claimed ~280 LOC of duplication; the actual net reduction was 127 lines. Honest accounting:

• ~150 lines of cross-method duplication genuinely eliminated (connect, isClosing, readLoop, waitRetry, startCloseWatcher, logPrefix, retry loop body).
• The per-variant handle logic ended up about the same size — the duplicated Meta[T]{Created, Updated, Index, Data} wrapping and the decode+OnError pattern still appear in both closures.
• New cost: function-pointer dispatch via s.handle, plus a runtime-attached closure that the reader has to trace back to which entry point wired it.

The actual bug class that motivated the entry — the cancel-suppression fix needing two copies in lockstep — was already addressed at the per-site level on #143 with regression coverage that catches a one-side-only regression. The Meta-wrapping duplication that remains is small and parallel-structured.

If consolidation comes back, go:generate (write the single version, generate the list version) is the lower-cost shape — no dispatch indirection, no closure capture.

[CBosch101]

Clean consolidation of the two near-identical subscription paths into one subscribeCore, with the variant-specific decode/delivery injected as a handle closure. Exported Subscribe[T]/SubscribeList[T] signatures are unchanged, so callers are unaffected. No blockers.

Verified

• Behavior invariants preserved line-for-line: single-object decode fatal-on-error; list parse fatal but per-item unmarshal fires OnError and continues; cache reassigned before the error check; retryCount=0 reset placed before OnMessage in both variants; cancel-suppression guard (isClosing()) intact at both connect() and readLoop() sites.
• meta import dropped cleanly — obj's type is now inferred from messages.Patch/PatchList; build and vet clean.
• run() loop matches both pre-refactor loops exactly; handle is always assigned before run(), no nil-call path.
• go test -race ./client/ . green, including the PR #143 cancel-suppression regression subtests.

The reduced test independence on the Subscribe-vs-SubscribeList axis (both wrappers now share connect/readLoop) is already documented in the updated test comments — agreed it's the right trade-off given the connect/readLoop axis still gates guard regressions.

---

🤖 Generated with Claude Code