plane: drop work_item.created echoes via external_source (PFB-27)

[hstern]

Summary

• handlePlaneWorkItemCreated now gates on WorkItem.ExternalSource starting with "forge:" and skips with reasonPlaneCreatedEchoExternalSource. The bridge stamps that prefix on every forge→plane create, so an inbound work_item.created carrying it is necessarily our own echo — mirroring it back duplicates the original forge issue.
• Non-bridge external_source values (e.g. operator-imported "github:..." work items) are unaffected — only the bridge prefix triggers the skip.
• AGENTS.md documents the underlying finding: Plane CE v1.3.1 strips HTML comments from description_html during ProseMirror sanitization, so the loop-break marker is non-functional on the description side. Plane preserves the marker in comment_html, so comment echoes still work via the marker.

Why

PFB-25 surfaced the duplicate path in production: a forge→plane create succeeded on the Plane side but the bridge errored decoding the REST response. The LRU was never populated (it's recorded after translator success). Plane's work_item.created echo arrived, the marker had been stripped by Plane during sanitization, and the bridge happily created a duplicate Forgejo issue.

PFB-25 closed the decode regression that triggered this specific instance, but the class of bug — any partial failure between Plane-side success and bridge-side success leaves the loop-break LRU empty and the marker is already gone — was open. This PR closes the structural gap.

Verified the sanitization claim against plane.stern.ca:

POST description_html: "<p>x</p>\n\n<!-- pfb:src=forge,evt=probe123 -->"
GET  description_html: "<div><p>x</p>\n\n</div>"
                       ^ marker dropped, body wrapped in <div>

POST comment_html:     "<p>my comment</p>\n\n<!-- pfb:src=forge,evt=c123 -->"
GET  comment_html:     "<p>my comment</p>\n\n<!-- pfb:src=forge,evt=c123 -->"
                       ^ marker preserved

Trade-offs considered

• Why not pre-record the LRU before POST returns (Option A from PFB-27 body)? The LRU keys are (side, deliveryID, targetObjID) — the forge-side delivery ID and plane-side delivery ID are unrelated, so pre-recording on one side doesn't help the other. Would require a separate cross-side keying scheme; complexity not justified given external_source covers the same threat.
• Why not validate external_source against the link's configured forge repo? A multi-link bridge with multiple forge repos still wants to skip ALL bridge-created work items; tying the skip to a specific link risks false-negatives if links change. The "forge:" prefix check is the right strictness.
• Why not also gate work_item.updated / deleted? Both currently skip with deferral reasons (no reverse lookup yet). When they're implemented, the same gate applies.

Test plan

• make race — all packages green
• make lint — 0 issues
• New regression tests: TestHandlePlaneWorkItem_CreatedEchoSkipsByExternalSource (bridge prefix → skip) and TestHandlePlaneWorkItem_CreatedNonBridgeExternalSourceProceeds (non-bridge prefix → proceed)
• Verified Plane sanitization behaviour against plane.stern.ca (probe commands captured in PR body)
• After merge: cut v0.1.4, pull :0 on the bridge host, confirm no duplicates on a clean forge→plane create even if response decode is artificially failed

🤖 Generated with Claude Code