[BUG] Legacy ledger synthesis labels .agents/ paths with target 'agents', which resolves to copilot (.github/)

[ahegyes]

Describe the bug

DeploymentLedgerCodec._legacy_locator() attributes a deployed path to a target by matching a path prefix against _LEGACY_TARGET_PREFIXES (apm_cli/core/deployment_ledger.py:29-39). Every entry in that map resolves back to a real target profile except .agents/, which maps to the string "agents".

"agents" is not a key in KNOWN_TARGETS, and normalize_target_name("agents") returns copilot, whose root_dir is .github — not .agents. APM's own deprecation warning states this explicitly:

'--target agents' is deprecated -- it maps to 'copilot' (.github/), not '.agents/'.

So when the deployment ledger is synthesized from the legacy deployed_files views (from_lockfile, deployment_ledger.py:180-192, taken whenever a lockfile carries no canonical deployments: block), every .agents/... row is written with a target that names a different target owning a different root.

.agents/ is additionally a shared root — codex, agent-skills, hermes, openclaw and antigravity all deploy there — so a single prefix cannot identify its owner. The other prefixes in the map (.claude/, .codex/, .github/, .cursor/, …) are each owned by exactly one target and round-trip correctly.

To Reproduce

Pure-function repro, no APM command run and no ~/.apm state touched:

import warnings; warnings.simplefilter("ignore")
from apm_cli.core.deployment_ledger import DeploymentLedgerCodec
from apm_cli.core.target_catalog import normalize_target_name
from apm_cli.integration.targets import KNOWN_TARGETS

for p in [".agents/skills/my-skill/SKILL.md",   # subject
          ".claude/skills/my-skill/SKILL.md",   # control
          ".codex/agents/my-agent.toml",        # control
          ".github/agents/my-agent.agent.md"]:  # control
    t = DeploymentLedgerCodec._legacy_locator(p).target
    norm = normalize_target_name(t)
    root = getattr(KNOWN_TARGETS.get(norm), "root_dir", "—")
    print(f"{p:38} target={t:8} normalizes={norm:8} in KNOWN_TARGETS={t in KNOWN_TARGETS!s:5} root={root}")

Output, identical on 0.28.0 and 0.29.0:

.agents/skills/my-skill/SKILL.md       target=agents   normalizes=copilot  in KNOWN_TARGETS=False root=.github   <-- mismatch
.claude/skills/my-skill/SKILL.md       target=claude   normalizes=claude   in KNOWN_TARGETS=True  root=.claude   ok
.codex/agents/my-agent.toml            target=codex    normalizes=codex    in KNOWN_TARGETS=True  root=.codex    ok
.github/agents/my-agent.agent.md       target=copilot  normalizes=copilot  in KNOWN_TARGETS=True  root=.github   ok

Three controls round-trip; only the subject does not, so the check discriminates.

Expected behavior

A synthesized legacy row should carry a target that actually owns the path, or the sentinel the function already has for "unattributable" ("legacy", the next(..., "legacy") default two lines below). It should not name a target whose deploy root does not contain the path.

Because .agents/ is shared by several targets, prefix matching cannot identify its owner from the path alone. Options that would both be sound: fall back to "legacy" for .agents/, or resolve the owner from the declared/active target set rather than the prefix.

Environment

• OS: macOS 26.6.2
• Python Version: 3.14.7
• APM Version: 0.28.0 and 0.29.0 (both reproduce; _LEGACY_TARGET_PREFIXES is byte-identical between them)

Additional context

Observed after upgrading 0.28.0 → 0.29.0 with a pre-existing lockfile, where a user-scope (-g) install produced a ledger with 233 rows labelled target: agents for .agents/skills/... paths, alongside correctly-labelled claude and codex rows for the same install. Re-running with the lockfile moved aside produced a ledger with no agents rows at all and correct per-target attribution.

I have not traced whether the mislabelled rows are what drove the subsequent reconciliation outcome, so I am reporting only the attribution defect above, which is self-contained and independently reproducible.