claimTags: whole-row id match makes the claim unwritable for most tasks on boards with a Depends column

[sshlg]

Summary

write_claim decides which table row is a task's claim by scanning the whole line for the
task id. Roadmaps that carry a Depends column repeat ids across rows, so most tasks match two or
more rows, acquire refuses to guess, and the durable claim tag is never written — while the
lease is granted normally.

The result is the state references/roadmap.md names as the thing to avoid:

A lease without a claim tag is invisible to anyone reading the repository.

Exclusion still holds. Visibility does not, and nothing about the run reports it as degraded.

Measured

Four-repository project (nicegram-business umbrella + three service submodules), agent-sync
1.3.5, claimTags: {"docs/ROADMAP.md": {"mode": "cell", "cell": -1, ...}}, boards shaped
| id | what | acceptance | depends | decisions | status |:

board	task ids	ids matching exactly one row	claim writable
account-factory	11	3	27%
account-session-connect	106	25	24%
nicegram-business-app	82	14	17%

The ids that do work are the ones nothing depends on — i.e. mostly leaf tasks. The more central
a task is, the less likely its claim is recorded.

Reproduction

| **AF-005** | Service skeleton | … | AF-001 | … | `todo` |
| **AF-006** | Schema and migrations | … | AF-005 | … | `todo` |

$ agent_sync.py acquire AF-005
  docs/ROADMAP.md: `AF-005` appears in 2 table rows — refusing to guess which one is the claim.
  Narrow the pattern or edit by hand
won AF-005 (run r-…, ttl 2700s)

Lease won, roadmap untouched. AF-005 is the second row's dependency, not its claim.

Cause

scripts/agent_sync.py, write_claim (~L1187):

hits = [i for i, l in enumerate(lines)
        if re.search(rf"(?<![A-Za-z0-9-]){re.escape(key)}(?![A-Za-z0-9-])", l)
        and self._row_cells(l) is not None]

The word-boundary match is right; its scope is the whole line. _row_cells is only used to
confirm the line is a row, never to say where in the row the id must appear.

The refusal message suggests "narrow the pattern", but claimTags keys are file globs — there is
no pattern that selects a row. The only available workarounds are editing the roadmap by hand
(which is what the lease exists to prevent) or dropping the Depends column.

Proposed fix

An optional idCell beside cell, 0-based with the same negative indexing, restricting the
search to one cell:

"apps/*/docs/ROADMAP.md": { "mode": "cell", "cell": -1, "idCell": 0,
                            "held": "{prev} (claimed: {holder})" }

def _matches(line):
    cells = self._row_cells(line)
    if cells is None:
        return False
    hay = line if id_cell is None else (cells[id_cell] if -len(cells) <= id_cell < len(cells) else "")
    return re.search(rf"(?<![A-Za-z0-9-]){re.escape(key)}(?![A-Za-z0-9-])", hay) is not None

Backward compatible: idCell unset keeps today's whole-row behaviour, so no existing config
changes meaning. check can validate it the same way it validates cell.

Two smaller things that would help independently of the above:

1. check could surface the ratio. It already refuses a claim-tag pattern with nothing to
   look for; a board where most ids are ambiguous is the same class of defect and is invisible
   until the first acquire.
2. acquire prints the refusal above the won line, so a | tail -3 hides it. Reporting the
   claim outcome after the lease result — or in the same line — would make it harder to miss.

Happy to open a PR for idCell if the shape looks right.