feat(zcoder): add GLM multi-model orchestrator with skill-forge lifecycle

[anthonykewl20]

What user problem does this solve

ZCode sessions routinely run every sub-task on the same model at the same effort, and skill edits ship on vibes. zcoder fixes both: it routes each decomposed sub-task to the right GLM engine with pinned thinking effort (vision / conditional-turbo / deep), always-on via three routing hooks; and it ships skill-forge, a measured skill lifecycle — deterministic Python-oracle grading into a Wilson-scored trials ledger, same-session PAIRED evidence, Pareto-frontier acceptance, reflective mutations, and a 7-gate interactive promotion flow.

What's included

• agents/ — glm-vision (GLM-5.3-Flash@max), glm-turbo (GLM-5.3-Flash@low, conditional lane), glm-main (GLM-5.3@max)
• skills/ — glm-orchestrator (routing protocol), skill-forge (lifecycle), laravel-dev + yaml-json-convert (first skills born from the lifecycle, each with scenario/oracle eval suites)
• commands/ — /orchestrate, /route, /skill-scan, /skill-evolve
• hooks/ — SessionStart / UserPromptSubmit / PreToolUse(Agent) routing injection with kill switches (~/.zcode/zcoder.off, GLM_ORCHESTRATOR_DISABLE=1)
• tests/ — the plugin's own portable, zero-token integrity suites (34-check OHI static, 29-check skill-forge static, 32-fixture smoke battery, coverage manifest, capability ledger) plus the perception calibration corpus
• EN + CN READMEs

How it was tested

• python3 scripts/validate.py → OK: 17 plugin(s) validated
• python3 scripts/build_dist.py → zcoder 0.1.4 artifact builds (119,478 bytes)
• git diff --check → clean
• Plugin self-suites run inside plugins/zcoder/ (installed-copy simulation): OHI static 34/34, skill-forge static 29/29, smoke battery 32/32, coverage manifest 66 files, 0 unaccounted
• The plugin has been dogfooded daily in ZCode (inline install): engine dispatch, hooks, commands, and two full skill-evolution rounds to promotion are exercised in the round log methodology (tests/OHI.md)

Version and marketplace registration

• New plugin zcoder v0.1.4, registered in marketplace.json with category developer-tools; manifest and marketplace entry share identical version and description_i18n (en + zh-CN).

Dependencies, network, permissions, side effects

• No network access and no MCP servers — the plugin makes zero outbound requests.
• Model dependency: GLM model IDs (glm-5.3, glm-5.3-flash) via ZCode's built-in Z.ai provider; README documents the one-line model: edit if a build's model picker uses different names.
• Command execution: one local bash hook script (hooks/inject-routing.sh) that reads its stdin payload and prints a routing directive; no file writes, no network, exits immediately when a kill switch is active.
• File writes: only when the user runs skill-forge rounds / OHI monitors — ledgers and logs under tests/ in the project where they run. No writes outside the project or the user-created ~/.zcode/zcoder.off flag.
• No credentials, telemetry, or analytics.

Licensing / provenance

• Plugin code: MIT.
• Methodology patterns (reflective mutation, size/growth caps, session mining, Pareto-frontier acceptance, staged mini-batch, trigger-surface evolution) ported from NousResearch/hermes-agent-self-evolution (MIT), attributed in the README. Its AGPL Darwinian Evolver was deliberately not ported; LLM-judge machinery was re-implemented as deterministic oracles.
• Perception corpus assets are generated by the shipped orchestrator-test/make_mockup*.py scripts (regenerable, no third-party content).

Checklist

• plugin name is unique and kebab-case (zcoder)
• required files and both language READMEs are present
• category is correct (developer-tools)
• no secrets, private data, or machine-specific paths (run ledgers/logs/state excluded; shipped suites pass on a clean copy)
• validate.py, build_dist.py, and git diff --check pass
• review feedback addressed (on this PR)