feat(agent): RSS-first production auth and workspace sessions

[fffonion]

Scope

Publishes the existing local RSS-first production-authentication stack without merging it into master.

• RSS owns provider/auth/refresh/selection/default/source/workspace behavior; Rust supplies bounded credential, OS, transport and resource capabilities.
• Includes the previously accepted local auth/provider and bundled-source work, the access-only Responses path, session/workspace admission, and packaged documentation corrections. The remote base predates these local commits, so this is the combined stack rather than a Task10-only diff.
• Workspace effects use retained confined authority. Admission/replay comparisons and atomic storage eligibility remain in RSS; persisted paths never grant authority.
• HTTP and Telegram workspace commands are followed by ordinary same-session runs with verified native file/process effects.
• The new workspace replay identity preserves the original optional selector and permits pure cached/durable replay before reopening workspace authority.

Current verification

Functional candidate: c7e931eb776f28985bad9bb590029445a17d0db5.
Current PR head: 1d48f40d6c6cb4396a5ef23805c804bc72cd829d, adding only the formatting repair requested by the first CI format check (one optional trailing argument comma removed, no argument/expression/assertion change). Local cargo fmt --all -- --check and a fresh current-artifact config_file_tests execution (21/21) passed after this repair. The initial remote CI failure and its subsequent rerun are kept distinct; no remote CI success is claimed here.

The subsequent CI run (https://github.com/rustscript-lang/rustscript-agent/actions/runs/38096048015) passed Format check, then failed Clippy (warnings denied) on clippy::collapsible_match at src/auth/authorization_output.rs:326. Its all-feature/all-target test step was skipped. This additional CI failure remains open; it must not be confused with the independently recorded lifecycle baseline failures below.

Final current-candidate execution used current-manifest Cargo JSON artifacts, exact feature/source checks, direct inner return codes, sequential builds and default-parallel test execution:

• 23 scoped gates: 22 passed; the full lifecycle target is separately qualified below.
• Documentation consistency: 9/9; documentation drift: 9/9; typed HTTP migration: 14/14.
• fmt, Clippy (-D warnings) and diff checks passed.
• Real RSS source discovery/corpus compilation passed; compile coverage is recorded separately from runtime acceptance.
• Narrow independent review closed original S2/S3/S5 and the two documentation compatibility regressions.
• New-format S1 cached/durable replay was covered. An additional independent compatibility probe used the exact supported-main 19603db0ff62527cb1ff1aa66f375ef98d7b06fc RSS storage writer with private business SQLite and reopened it using the current consumer: the same request returned the original run, with no new messages/events; adding a selector was rejected.

Known failing target

Full run_lifecycle_tests returned RC 101: 32 passed, 2 failed:

• hanging_http_adapter_stop_cancels
• test_sources_do_not_hardcode_sibling_lease_temp_roots

Both failures were independently classified against the prior base; they remain recorded and are not reported as target success. Original assertions, timing limits, process cleanup and admission bounds were retained. A historical transient pending-count observation race is also retained as an existing advisory; later passing runs do not erase it.

Compatibility qualification

The narrow reviewer raised a conditional S1 concern about reusing the old workspace-scoped idempotency format from unaccepted candidate b8d46bd75db108bfba911f8e49a0c03ac02e5533. That prerelease format has not been integrated into master or deployed in this execution and is not treated as a supported deployed format. Supported-main compatibility was exercised separately as described above. No legacy native business-hash implementation was introduced.

Remaining work

This is a draft integration PR, not a completed release:

• Task11: production approval decisions and execution gating.
• Task12: production compaction wiring.
• Task13: remaining migration/release integration and final full-suite acceptance.
• Task8: full live refresh validation is deferred; the accepted live path used access tokens only.

No auto-merge is enabled. The PR does not claim all debug/release suites or whole-plan completion.

[fffonion]

Superseded by #5.

The complete commit lineage of this PR is already contained in #5, which also carries the subsequent production integration fixes and current validation. Closing this historical draft in favor of #5 to avoid duplicate review. Follow #5 for the remaining exact-head CI acceptance.

This closure does not merge #4 into master.

[fffonion]

PR #4 implemented-feature coverage was checked against the production-only source and the extracted SSE framer. The successor production candidate is #6 (c65ac13), with full Debug coverage and exact-head CI passed. Two local Release fixture cancellations remain unresolved, so #6 is not merged. Current tool-plugin/effects work is separate in Draft #7. These successors are native stack #8 (#6 → #7); combined PR #5 is now closed without merging.

[fffonion]

Production successor #6 is now merged into master (3b6fcc2a7872997799d07a25a7bd91deb51b17b2). PR #4 implemented-feature coverage was verified through the production source and current runtime/CI gates. Current tool-plugin/effects work is preserved separately in Draft #7, now based on master after native restacking. Original combined PR #5 remains closed without merging; the old full source is preserved. Historical SQLite cancellations remain recorded without an unsupported root-cause claim.