refactor: resolve goconst findings via constants and parser unification

[kazz187]

Summary

golangci-lint 2.13.1 upgraded goconst 1.8.2 → 1.11.0, which widened repeated-literal counting from file scope to package scope. These are legitimate detections rather than a lapsed config, so they are resolved as a real refactor instead of being suppressed.

• .golangci.yml — goconst.ignore-tests: true (test fixture duplication is desirable, and goconst counts across the package so exclusions.rules cannot substitute) plus ignore-string-values: ['^(?:true|false)$'] for the stringly-typed metadata booleans. Thresholds unchanged; ignore-map-keys deliberately not used.
• internal/claudemd (new) — unifies the four duplicated frontmatter parsers (skill/agent × file/string) behind ParseSkill / ParseAgent. Net −878 lines across the four call sites.
• internal/eventbus/metadata.go (new) — 19 event metadata keys as shared constants, replacing literals in 16 producer files and 4 reader sites.
• cmd/taskguild-agent/constants.go (new) — tool names, interaction option values, hook decisions, settings.json keys, task-log metadata keys and content block fields.
• pkg/clog / internal/template / internal/project — log attribute keys, template entity types, seeded workflow status names and model.

Also folded in two pre-existing duplications flagged during the work: single_command_permission_cache.go now reuses singlecommandpermission.TypeCommand/.TypeRedirect, and the twice-declared startMarker/endMarker block in directive.go is hoisted to package scope.

Results

	before	after
goconst	221	0
total lint findings	2126	1873 (−253)
linters above baseline	—	none

Notes for review

• Measure with --uniq-by-line=false. golangci-lint rounds to one finding per line by default, which understated the real count (e.g. prompt.go:125 carries both "type" and "text" but reports one). The true non-test count was 99 across 48 distinct literals, not 90.
• No string value changed. Every literal→constant substitution was machine-verified by expanding constant identifiers in each added line back to their declared values and comparing against the removed line — 222/222 identical. The metadata keys are a wire contract with the frontend (event-stream.ts, WorktreeList.tsx), so the values are frozen.
• Deliberately left alone: slog attribute-name literals (different domain from event metadata, and excluded by goconst's default exclude-types: [Call]), task-log metadata maps (full_text / result_type / status_id / source), and display strings such as "**Tool:** `Bash`\n".
• internal/schedule/server_test.go is an unrelated golangci-lint fmt fix of the pre-existing gofumpt finding.

Behaviour change to check

The string-based agent parser in internal/agentmanager/agent_conflict.go previously used strings.CutPrefix and supported neither block scalars, list-form tools:/disallowedTools:, indented keys, nor context. After unification it handles all of them. This is strictly more correct, and the new package has 31 unit tests covering it — but the ResolveAgentConflict AGENT-choice path has not been exercised against real data.

Test plan

• go build ./...
• go test ./... — all packages pass; the 13 existing parser tests in internal/skill / internal/agent pass unchanged, which is the regression guard for the unification
• go tool golangci-lint run --default=none --enable=goconst --max-issues-per-linter 0 --max-same-issues 0 --uniq-by-line=false ./... → 0 issues
• Full lint compared against baseline — no linter regressed
• internal/claudemd lints clean under default: all
• Exercise agent/skill conflict resolution (AGENT choice) against a real .claude/agents/*.md with block scalars