docs: attribute the write-ahead login work to 0.0.14 (WALM-649)

docs/mcp/changelog.mdx

@@ -35,6 +35,10 @@ answer: >-
 
 Unreleased package version. Plugin launchers in this tree pin `@mysten-incubation/memwal-mcp@0.0.14-dev.0` (the published `dev` dist-tag) until `0.0.14` exists on npm. Unpin `.mcp.json` / `.cursor-mcp.json` / `.codex-mcp.json` to `@0.0.14` in the same release that publishes it, or plugin installs stay on the dev build.
 
+### Added
+
+- Opt-in Streamable HTTP transport for the stdio bridge. Set `MEMWAL_MCP_TRANSPORT=http` (aliases `streamable`, `streamable-http`) to dial the relayer's single `/api/mcp` endpoint, where a call is answered on the same request instead of being split across a POST and an SSE stream, so there is no idle watchdog. The default stays `sse` and unrecognised values fall back to it. Reconnect replay is not transport-aware yet: on Streamable a disconnect mid-send can look sent, so a replayed write may duplicate.
+
 ### Fixed
 
 - The cold-start tool list no longer advertises a tool the relayer may not serve. The bridge ships on npm and updates itself while a relayer ships per environment, so 0.0.14-dev.0 dialled prod and staging still on 0.0.13: cold start named `memwal_remember_status`, which neither registers, and the pending-write wording sent the agent to go call it — one live run spent 90.67s there before erroring. Cold start is now a floor rather than a forecast (`BASELINE_RELAYER_TOOLS`): it carries only what the oldest supported relayer serves and its descriptions name nothing outside it, while newer tools still reach the client a beat later on the relayer's own `tools/list`. Initialize `instructions` use the same floor (they used to name `memwal_remember_status` before any `tools/list`). And a call for a tool outside that floor — including during the cold-start window, before any upstream `tools/list` has been seen — is now answered locally and at once — naming the tools that do exist and saying plainly that nothing ran — instead of being forwarded into a wait that only ends at the orphan deadline. (#928)
@@ -43,22 +47,22 @@ Unreleased package version. Plugin launchers in this tree pin `@mysten-incubatio
 - `memwal_remember` sends a content-derived idempotency key, so the retry its own timeout message invites really does attach to the job already in flight instead of storing a second paid copy. The key is computed by the tool rather than relied on from the SDK, whose published build mints a random UUID per client instance.
 - `memwal_remember_status` accepts `job_ids` to settle a whole batch in one call, and reports a mixed batch honestly — a still-uploading row no longer renders the poll timeout as `error=`, which read as a failed write. Settling in one request also matters against the rate limit: 20 ids cost one request, not twenty.
 - The bridge's cold-start tool list no longer disagrees with the sidecar's. `memwal_remember_status` advertised only `job_id`, required, under `additionalProperties: false`, so the batch call the tools themselves instruct was rejected until `tools/list_changed` arrived; the `waitMs` ceiling advertised 60000 after the sidecar lowered it to 45000, which came back as an MCP validation error; and `memwal_remember_bulk` still carried its pre-queue description. Tests now pin the parts an agent acts on.
+- Persist the delegate keypair before sign-in hands the URL to the browser, and reclaim it on the next start. The browser's onchain `add_delegate_key` costs gas and is irreversible, and it happens before the callback that saves the private half, so a client that died in that window destroyed the only copy of a key the user had already paid for and left an orphaned registration nobody could use. (#793) Signing out discards the pending record along with the credentials, a second sign-in against the same relayer reuses the stranded key rather than minting over it, and a sign-in that cannot write the record fails instead of publishing a URL it cannot back. Reclaiming works on Mainnet; Testnet requires an account-id hint the recovering client does not have.
 
 ### Changed
 
 - `memwal_remember` / `memwal_remember_bulk` return at accept by default (`MEMWAL_MCP_REMEMBER_WAIT_MS=0`, ~1s, `job_id`). The Walrus write continues in the background; do not treat that reply as stored. Set `MEMWAL_MCP_REMEMBER_WAIT_MS=90000` to restore wait-for-`blob_id` (90s ceiling). Do not use a value between 0 and the real completion time — that pays the wait and still returns pending. The MCP TypeScript SDK's default tools/call timeout is 60s, so a 90s wait loses on hosts that do not raise it.
 
 ## 0.0.13
 
-This release stops a write whose reply was lost from being reported as safe to retry — repeating one can store a second paid copy — answers tool calls with an auth error pointing at `memwal_login` when the relayer rejects the saved delegate key, saves the delegate keypair before sign-in hands a URL to the browser and reclaims it on the next start, writes the credentials file through a fresh `0600` file that it renames into place, confirms a completed sign-in and keeps the bridge reading stdin afterwards, warns on unrecognised command-line options instead of ignoring them, documents the network presets in `--help`, names the relayer in `memwal_health`, reports restore `failed` counts when truncation is a transient download or embed blip, and pins plugin launch configs (`.mcp.json`, Cursor/Codex copies, and the Codex fallback installer) so npx cannot keep a cached 0.0.5.
+This release stops a write whose reply was lost from being reported as safe to retry — repeating one can store a second paid copy — answers tool calls with an auth error pointing at `memwal_login` when the relayer rejects the saved delegate key, writes the credentials file through a fresh `0600` file that it renames into place, confirms a completed sign-in and keeps the bridge reading stdin afterwards, warns on unrecognised command-line options instead of ignoring them, documents the network presets in `--help`, names the relayer in `memwal_health`, reports restore `failed` counts when truncation is a transient download or embed blip, and pins plugin launch configs (`.mcp.json`, Cursor/Codex copies, and the Codex fallback installer) so npx cannot keep a cached 0.0.5.
 
 ### Fixed
 
 - Stop telling the user to retry a write whose reply was lost. A `memwal_remember`, `memwal_remember_bulk` or `memwal_analyze` that was POSTed and then timed out came back as "the connection to the relayer dropped before the result came back. Please retry." — but the relayer answers those with HTTP 202 and finishes the work in a durable queue, so a client-side deadline cancels nothing and the write may already have landed. `/api/remember/bulk` carries no idempotency key, unlike the single path, so following that advice stores a second paid copy that `recall` then hides behind the first. A sent write now says it may have completed, that the timeout did not undo it, and to check with `memwal_recall` before re-saving. A sent read still says plainly that retrying is safe. (WALM-618 follow-up)
 - Answer tool calls with an auth error when the relayer rejects the saved delegate key, instead of parking them until the call deadline. A 401 on the SSE handshake was treated like any other connect failure, so the bridge retried a key that could never be accepted while the queued `memwal_recall` waited out the orphan sweeper — up to four minutes — and then came back as "the connection to the relayer dropped, please retry", advice that cannot work. The bridge now names the rejection and points at `memwal_login`, whether the key is rejected at startup or revoked mid-session, and refuses later requests immediately while it stays rejected. Any accepted handshake resumes normal buffering, so both a re-login and a transient WAF or rate-limit 401 recover on their own. Credentials are still never wiped automatically. (#365, WALM-602)
 - Back off when the relayer refuses the SSE handshake with HTTP 429 instead of retrying ~500ms later. The bridge now honours a `Retry-After` (clamped to 60s) and falls back to a 5s floor when the header is absent — the `ip_active_cap` shape — and carries the deadline across reconnect attempts. It also prints one stderr line saying this is a rate limit rather than a bad config or bad credentials, so a throttled bridge no longer reads as a broken one. (WALM-386)
 - Answer a request that is still buffered while the handshake keeps failing, instead of holding it for the full call timeout and then blaming a dropped reply. A request that was never sent cannot have executed, so after 90s of consecutive handshake failures the bridge fails it with the reason the handshake actually gave — rather than parking it four minutes and returning "the connection to the relayer dropped, please retry", which named the wrong layer and invited a retry of a `remember` that had never left the process. The full deadline still applies while the handshake is healthy. Override with `MEMWAL_MCP_STALLED_HANDSHAKE_MS`. (WALM-618)
-- Persist the delegate keypair before sign-in hands the URL to the browser, and reclaim it on the next start. The browser's onchain `add_delegate_key` costs gas and is irreversible, and it happens before the callback that saves the private half, so a client that died in that window destroyed the only copy of a key the user had already paid for and left an orphaned registration nobody could use. (#793) Signing out discards the pending record along with the credentials, a second sign-in against the same relayer reuses the stranded key rather than minting over it, and a sign-in that cannot write the record fails instead of publishing a URL it cannot back. Reclaiming works on Mainnet; Testnet requires an account-id hint the recovering client does not have.
 - `memwal_restore` reports `failed` and retries the same page when `truncated` is a download/embed blip (`restored=0` and `skipped+failed < total`), instead of always telling the agent to raise `limit` (WALM-480).
 - Unrecognised options now warn on stderr and in the structured log (`cli.unrecognised_arg`) instead of being dropped in silence, so a typo'd `--namesapce work` no longer writes to the default namespace with nothing to say it had. The warning names the option key only, keeping a mistyped value-taking flag (`--tokenn=hunter2`) from putting the secret on stderr, and it warns rather than exits so an option from a newer config cannot brick the server. (#630)
 - `--help` now lists the network presets (`--prod`, `--dev`, `--staging`, `--local`) with the relayer and web URLs each resolves to, rendered from the preset table rather than retyped so a new preset cannot ship undocumented the way `--prod` did. The `--label` default is corrected to "MCP Client", which is what the code actually falls back to. (#630)

docs/sdk/changelog.mdx

@@ -28,7 +28,7 @@ questions:
   - When was bulk remember added to the Walrus Memory SDK?
   - What security improvements have been made to the MemWal SDK?
 answer: >-
-  The latest TypeScript SDK release is 0.1.7. Request bodies are hashed with `@noble/hashes` rather than WebCrypto-or-`node:crypto`, so the SDK no longer imports a Node builtin that Vite silently externalises into a runtime crash in the browser, and it declares a Node 20 floor. `restore()` results include `failed` (required like `truncated`; SDK defaults omitted to `0`) for permanent decrypt/UTF-8 failures instead of folding them into `skipped`. Empty-body 401s use the AUTH_REJECTED troubleshooting message instead of telling callers to run `memwal_login`. Account and manual PTBs use typed `tx.pure` helpers so they work with modern `@mysten/sui`. An explicit `sort` on `recall()`, `"relevance"` included, now makes the relayer ignore `scoringWeights`; weights re-rank only when `sort` is omitted. 0.1.6 added optional `created_at` on `recall()` results, plus `sort` and `scoringWeights` on `RecallOptions`, and reports HTTP 503 as a retryable upstream outage instead of a sign-in failure.
+  The latest TypeScript SDK release is 0.1.8. Request bodies are hashed with `@noble/hashes` rather than WebCrypto-or-`node:crypto`, so the SDK no longer imports a Node builtin that Vite silently externalises into a runtime crash in the browser, and it declares a Node 20 floor. `restore()` results include `failed` (required like `truncated`; SDK defaults omitted to `0`) for permanent decrypt/UTF-8 failures instead of folding them into `skipped`. Empty-body 401s use the AUTH_REJECTED troubleshooting message instead of telling callers to run `memwal_login`. Account and manual PTBs use typed `tx.pure` helpers so they work with modern `@mysten/sui`. An explicit `sort` on `recall()`, `"relevance"` included, now makes the relayer ignore `scoringWeights`; weights re-rank only when `sort` is omitted. 0.1.6 added optional `created_at` on `recall()` results, plus `sort` and `scoringWeights` on `RecallOptions`, and reports HTTP 503 as a retryable upstream outage instead of a sign-in failure.
 ---
 
 ## 0.1.8